Oona Rainio
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Oona Rainio has a total of 55 co-authors. Here are the top co-authors by number of shared publications:
- Riku Klén
Turku PET Centre, University of Turku, Turku, Finland
30 shared publications - Matti Vuorinen
Department of Mathematics and Statistics, University of Turku, Turku, Finland
27 shared publications - Jukka Kemppainen
Turku PET Centre, University of Turku and Turku University Hospital, Turku, Finland
8 shared publications - Juhani Knuuti
Turku PET Centre, University of Turku, Turku, Finland
7 shared publications - Mohamed M. S. Nasser
Department of Mathematics, Statistics, and Physics, Wichita State University, Wichita, USA
7 shared publications - Joonas Liedes
Turku PET Centre, University of Turku and Turku University Hospital, Turku, Finland
5 shared publications
Latest Publications
- Open Access
- 16 October 2026
- Published by Springer Nature in Journal of Imaging Informatics in Medicine
- p. 1-11
Investigation on Automatic Classification of Bone, Lymph Node, and Local Lesions in CT Images of Prostate Cancer Patients
Accurate assessment of different types of prostate cancer (PCa) lesions is a crucial part of treatment planning. Following a computed tomography (CT) or a positron emission tomography (PET)/CT imaging scan, a significant amount of time is required from the clinicians to detect lesions and characterize them by their anatomical locations. To facilitate the development of new image analysis applications, we investigated the potential of automated classification of PCa lesions between bone metastases, lymph node metastases, local lesions, and other PCa lesions in CT images via machine learning (ML) and deep learning models. We annotated a total of 4949 PCa lesions by their type in a private dataset of $$^{18}$$ F-prostate-specific membrane antigen (PSMA) PET/CT images of 532 PCa patients and an external test set of publicly available PSMA PET/CT images from 30 patients. Using information of coordinates of the detected lesions, we automatically extracted 15 different features from the CT and the results of the automated 3D anatomical segmentation tool TotalSegmentator for each lesion, and trained several traditional ML models to classify lesions by their anatomical type. Additionally, we experimented with convolutional neural networks (CNNs) of different dimensionalities with CT and TotalSegmentator inputs. According to our results, the most accurate model was XGBoost with an overall accuracy of 96.0%. It identified bone metastases, lymph node metastases, and local lesions with sensitivities of 97.9±1.2%, 95.2±1.3%, and 93.1±2.8% in the five-fold cross-validation experiments of the primary data and 96.3%, 94.0%, and 80.0% on the external test set, respectively. Conversely, the CNNs underperformed the best ML models in distinguishing local lesions. To conclude, a simple ML model can accurately classify PCa lesions between bone, lymph node, and local lesions based on CT images, lesion locations, and automated anatomical segmentation data, but more research is required for distinguishing PCa lesions outside these categories.- 16 October 2026
- Published by Springer Nature in Journal of Imaging Informatics in Medicine
- p. 1-11
Open AccessInvestigation on Automatic Classification of Bone, Lymph Node, and Local Lesions in CT Images of Prostate Cancer Patients
Accurate assessment of different types of prostate cancer (PCa) lesions is a crucial part of treatment planning. Following a computed tomography (CT) or a positron emission tomography (PET)/CT imaging scan, a significant amount of time is required from the clinicians to detect lesions and characterize them by their anatomical locations. To facilitate the development of new image analysis applications, we investigated the potential of automated classification of PCa lesions between bone metastases, lymph node metastases, local lesions, and other PCa lesions in CT images via machine learning (ML) and deep learning models. We annotated a total of 4949 PCa lesions by their type in a private dataset of $$^{18}$$ F-prostate-specific membrane antigen (PSMA) PET/CT images of 532 PCa patients and an external test set of publicly available PSMA PET/CT images from 30 patients. Using information of coordinates of the detected lesions, we automatically extracted 15 different features from the CT and the results of the automated 3D anatomical segmentation tool TotalSegmentator for each lesion, and trained several traditional ML models to classify lesions by their anatomical type. Additionally, we experimented with convolutional neural networks (CNNs) of different dimensionalities with CT and TotalSegmentator inputs. According to our results, the most accurate model was XGBoost with an overall accuracy of 96.0%. It identified bone metastases, lymph node metastases, and local lesions with sensitivities of 97.9±1.2%, 95.2±1.3%, and 93.1±2.8% in the five-fold cross-validation experiments of the primary data and 96.3%, 94.0%, and 80.0% on the external test set, respectively. Conversely, the CNNs underperformed the best ML models in distinguishing local lesions. To conclude, a simple ML model can accurately classify PCa lesions between bone, lymph node, and local lesions based on CT images, lesion locations, and automated anatomical segmentation data, but more research is required for distinguishing PCa lesions outside these categories. - Open Access
- 8 September 2026
- Published by Springer Nature in Physical and Engineering Sciences in Medicine
- p. 1-9
Deep learning for automatic detection of prostatectomy in CT images of prostate cancer patients
Extracting anatomic reference from computed tomography (CT) images is a crucial step in fully automated image analysis solutions for CT and hybrid imaging of prostate cancer. Several deep learning-based applications can be used to perform segmentation of different anatomical structures in CT, but they might produce a false prostate segment for post-treatment scans of patients treated with prostatectomy. In this study, our aim was to both systematically assess the performance of a state-of-the-art CT segmentation tool, TotalSegmentator, for post-prostatectomy patients and to investigate the potential of using a convolutional neural network (CNN) to automatically detect prostatectomy from CT images. We collected a dataset of CT images from 542 patients, 269 of which were treated with robotic-assisted laparoscopic radical prostatectomy (RP), 194 of which were treated with radiation and/or androgen deprivation therapy only, and 79 of which were treatment-naive. We used TotalSegmentator to perform multi-organ segmentation for all the patients, computed the volumes of the greatest connected components of the prostate segments, and studied the use of a cut-off threshold for the resulting volumes to detect RP with five-fold cross-validation. Additionally, we trained and evaluated a light-weight CNN for classifying patients treated with and without RP. According to our results, TotalSegmentator produced a false prostate segment for 98.5% patients treated with RP. The prostate segments were significantly smaller in the RP cohort than the other two cohorts (12.1 ± 5.8 cm $$^3$$ vs. 22.8 ± 8.9 cm $$^3$$ and 22.4 ± 10.9 cm $$^3$$ , p -values: 1.3e−17 and 2.1e−17). The use of cut-off thresholds for TotalSegmentator’s prostate volumes resulted in RP detection accuracy of 77.5 ± 1.3%, sensitivity of 83.0 ± 3.9%, specificity of 72.2 ± 3.5%, precision of 74.4 ± 3.1%, and area under receiver operating characteristic curve (auROC) of 84.1 ± 2.3%. Conversely, our proposed CNN approach detected RP with accuracy of 86.5 ± 4.2%, sensitivity of 78.9 ± 12.5%, specificity of 93.5 ± 3.9%, precision of 92.7 ± 3.9%, and auROC of 95.1 ± 1.7%, outperforming the threshold approach in a statistically significant way according to DeLong’s tests (4 out of 5 p -values<0.045). To conclude, TotalSegmentator systematically produces false prostate segments with patients treated with RP but, by utilizing a CNN, RP can be detected automatically based on CT data only. This enables automated correction of TotalSegmentator masks. If developed further, our CNN method could provide a stand-alone application for identifying presence of prostatectomy efficiently without the need to consult treatment records and, consequently, assist in developing fully-automated image analysis tools.- 8 September 2026
- Published by Springer Nature in Physical and Engineering Sciences in Medicine
- p. 1-9
Open AccessDeep learning for automatic detection of prostatectomy in CT images of prostate cancer patients
Extracting anatomic reference from computed tomography (CT) images is a crucial step in fully automated image analysis solutions for CT and hybrid imaging of prostate cancer. Several deep learning-based applications can be used to perform segmentation of different anatomical structures in CT, but they might produce a false prostate segment for post-treatment scans of patients treated with prostatectomy. In this study, our aim was to both systematically assess the performance of a state-of-the-art CT segmentation tool, TotalSegmentator, for post-prostatectomy patients and to investigate the potential of using a convolutional neural network (CNN) to automatically detect prostatectomy from CT images. We collected a dataset of CT images from 542 patients, 269 of which were treated with robotic-assisted laparoscopic radical prostatectomy (RP), 194 of which were treated with radiation and/or androgen deprivation therapy only, and 79 of which were treatment-naive. We used TotalSegmentator to perform multi-organ segmentation for all the patients, computed the volumes of the greatest connected components of the prostate segments, and studied the use of a cut-off threshold for the resulting volumes to detect RP with five-fold cross-validation. Additionally, we trained and evaluated a light-weight CNN for classifying patients treated with and without RP. According to our results, TotalSegmentator produced a false prostate segment for 98.5% patients treated with RP. The prostate segments were significantly smaller in the RP cohort than the other two cohorts (12.1 ± 5.8 cm $$^3$$ vs. 22.8 ± 8.9 cm $$^3$$ and 22.4 ± 10.9 cm $$^3$$ , p -values: 1.3e−17 and 2.1e−17). The use of cut-off thresholds for TotalSegmentator’s prostate volumes resulted in RP detection accuracy of 77.5 ± 1.3%, sensitivity of 83.0 ± 3.9%, specificity of 72.2 ± 3.5%, precision of 74.4 ± 3.1%, and area under receiver operating characteristic curve (auROC) of 84.1 ± 2.3%. Conversely, our proposed CNN approach detected RP with accuracy of 86.5 ± 4.2%, sensitivity of 78.9 ± 12.5%, specificity of 93.5 ± 3.9%, precision of 92.7 ± 3.9%, and auROC of 95.1 ± 1.7%, outperforming the threshold approach in a statistically significant way according to DeLong’s tests (4 out of 5 p -values<0.045). To conclude, TotalSegmentator systematically produces false prostate segments with patients treated with RP but, by utilizing a CNN, RP can be detected automatically based on CT data only. This enables automated correction of TotalSegmentator masks. If developed further, our CNN method could provide a stand-alone application for identifying presence of prostatectomy efficiently without the need to consult treatment records and, consequently, assist in developing fully-automated image analysis tools. - Open Access
- 26 August 2026
- Published by Springer Nature in Journal of Medical and Biological Engineering
- Vol. 46 (4) , 513-522
Image-derived Input Functions for Cerebral Blood Flow Quantification from Dynamic 15O-water PET Images of Humans
Purpose: Cerebral blood flow (CBF) quantification from dynamic $$^{15}$$ O-water positron emission tomography (PET) images via compartmental modeling requires measuring the tracer concentration in the arterial blood with respect to time. The existing methods for this include arterial blood sampling or extracting an image-derived input function (IDIF) from carotid arteries, neither of which is uncomplicated. However, new long axial field of view PET scanners have recently enabled the IDIF extraction from sites outside the head and neck area, such as the aorta, and new automatic segmentation tools allow an effortless volume of interest (VOI) definition. Here, our aim is to investigate the potential of the IDIFs from automatically generated aorta and common carotid artery (CCA) VOIs. Methods: We compare the aorta and CCA IDIFs extracted from TotalSegmentator-defined VOIs with and without accounting for internal dispersion. We also introduce a new hybrid IDIF approach based on the use of CCA IDIF to adjust dispersion in the aorta IDIF. We systematically compute the CBF estimates with these five IDIFs from dynamic total-body $$^{15}$$ O-water PET images of 100 human patients. Results: Compared to CCA IDIFs, the aorta IDIFs produced better model fit and less patient-wise variation, had no outlier values, and resulted in CBF estimates more consistent to earlier literature. Adjusting the dispersion significantly improved the model fit and affected the numerical values of the CBF estimates for both aorta and the CCA IDIFs. Conclusions: The use of an IDIF extracted from an automatically generated aorta VOI in dynamic total-body $$^{15}$$ O-water PET images might offer an effective way of CBF quantification. More research is warranted for its validation. Based on our results, the use of automatically generated CCA VOIs is not recommended for this purpose.- 26 August 2026
- Published by Springer Nature in Journal of Medical and Biological Engineering
- Vol. 46 (4) , 513-522
Open AccessImage-derived Input Functions for Cerebral Blood Flow Quantification from Dynamic 15O-water PET Images of Humans
Purpose: Cerebral blood flow (CBF) quantification from dynamic $$^{15}$$ O-water positron emission tomography (PET) images via compartmental modeling requires measuring the tracer concentration in the arterial blood with respect to time. The existing methods for this include arterial blood sampling or extracting an image-derived input function (IDIF) from carotid arteries, neither of which is uncomplicated. However, new long axial field of view PET scanners have recently enabled the IDIF extraction from sites outside the head and neck area, such as the aorta, and new automatic segmentation tools allow an effortless volume of interest (VOI) definition. Here, our aim is to investigate the potential of the IDIFs from automatically generated aorta and common carotid artery (CCA) VOIs. Methods: We compare the aorta and CCA IDIFs extracted from TotalSegmentator-defined VOIs with and without accounting for internal dispersion. We also introduce a new hybrid IDIF approach based on the use of CCA IDIF to adjust dispersion in the aorta IDIF. We systematically compute the CBF estimates with these five IDIFs from dynamic total-body $$^{15}$$ O-water PET images of 100 human patients. Results: Compared to CCA IDIFs, the aorta IDIFs produced better model fit and less patient-wise variation, had no outlier values, and resulted in CBF estimates more consistent to earlier literature. Adjusting the dispersion significantly improved the model fit and affected the numerical values of the CBF estimates for both aorta and the CCA IDIFs. Conclusions: The use of an IDIF extracted from an automatically generated aorta VOI in dynamic total-body $$^{15}$$ O-water PET images might offer an effective way of CBF quantification. More research is warranted for its validation. Based on our results, the use of automatically generated CCA VOIs is not recommended for this purpose. - Open Access
- 30 March 2026
- Published by Springer Nature in Journal of Imaging Informatics in Medicine
Deep Learning for Medical Ultrasound Image Segmentation: A Systematic Review of the Current Research
Deep learning (DL) has enabled automated segmentation of ultrasound images, and due to the rapid development of DL models, we want to offer a comprehensive overview of the current state of research. Following PRISMA 2020 guidelines, we systematically selected and analyzed 296 recent scientific articles on DL-based ultrasound segmentation from the PubMed database. According to our results, the most common targets of DL-based ultrasound segmentation are breast tumors, organs, and cardiovascular structures. Other major application categories include orthopedics, thyroid nodules, obstetrics-gynecology, and oncology in general. Convolutional neural networks (CNNs) and especially U-shaped architectures have preserved their popularity, even though vision transformers (ViTs), CNN/ViT hybrids, and segment anything models have also become well-established within a few years of their release. The newer models are given significantly more data, but no association between the method type and the reported values of the evaluation metrics can be detected across several studies. Most common limitations of the current research include a lack of information on computational requirements and issues related to model performance evaluation. DL-based ultrasound segmentation is a quickly developing field, supported by increased use of ultrasound imaging, new public datasets, and methodological advancements.
- 30 March 2026
- Published by Springer Nature in Journal of Imaging Informatics in Medicine
Open AccessDeep Learning for Medical Ultrasound Image Segmentation: A Systematic Review of the Current Research
Deep learning (DL) has enabled automated segmentation of ultrasound images, and due to the rapid development of DL models, we want to offer a comprehensive overview of the current state of research. Following PRISMA 2020 guidelines, we systematically selected and analyzed 296 recent scientific articles on DL-based ultrasound segmentation from the PubMed database. According to our results, the most common targets of DL-based ultrasound segmentation are breast tumors, organs, and cardiovascular structures. Other major application categories include orthopedics, thyroid nodules, obstetrics-gynecology, and oncology in general. Convolutional neural networks (CNNs) and especially U-shaped architectures have preserved their popularity, even though vision transformers (ViTs), CNN/ViT hybrids, and segment anything models have also become well-established within a few years of their release. The newer models are given significantly more data, but no association between the method type and the reported values of the evaluation metrics can be detected across several studies. Most common limitations of the current research include a lack of information on computational requirements and issues related to model performance evaluation. DL-based ultrasound segmentation is a quickly developing field, supported by increased use of ultrasound imaging, new public datasets, and methodological advancements.
- Open Access
- 23 February 2026
- Published by Springer Nature in European Radiology Experimental
- Vol. 10 (1) , 15
Acute deep neck infection MRI: deep learning segmentation and clinical relevance of retropharyngeal edema volume
Objective: Retropharyngeal edema (RPE) on MRI in patients with acute neck infection is associated with disease severity. We explored the potential role of RPE volume as a quantitative marker and developed a convolutional neural network (CNN) for automated RPE volume segmentation. Materials and methods: Volumes of RPE were manually segmented from T2-weighted fat-suppressed Dixon magnetic resonance (MR) images from 244 patients. These volumes were correlated with clinical variables, such as the need for intensive care unit (ICU) admissions, C-reactive protein (CRP) levels, maximal abscess diameter, and length of hospital stay (LOS). Manually segmented masks were used to train a CNN. Results: Patients who required ICU admission had significantly higher RPE volumes than those who did not, and RPE volume outperformed the binary RPE (presence/absence) in classification analysis of ICU admissions. Furthermore, RPE volume correlated positively with LOS, CRP, and maximal abscess diameter. At the slice level, the deep learning (DL)-based model achieved its highest area under the receiver operating characteristic curve (AUROC) in sagittal slices (98.2%) and its highest Dice similarity coefficient in axial slices (0.534). Conclusion: RPE volume is a promising quantitative imaging biomarker associated with relevant clinical outcomes in acute neck infections. Our DL-based model enables automated quantification of RPE volume. Relevance statement: RPE volume provides clinically meaningful information in acute neck infections, outperforming binary classification in predicting disease severity and correlating with key clinical outcomes. Automated DL-based segmentation accurately locates the RPE and provides a moderate quantitative measurement of RPE volume, supporting its potential as a clinical imaging biomarker. Key Points: RPE volume correlated with markers of severe illness and outperformed binary RPE classification. We developed a DL-based algorithm for slice-wise classification and automatic segmentation of RPE. The classification model achieved excellent performance, while segmentation yielded modest Dice similarity coefficients consistent with prior imaging-based tumor segmentation algorithms.
- 23 February 2026
- Published by Springer Nature in European Radiology Experimental
- Vol. 10 (1) , 15
Open AccessAcute deep neck infection MRI: deep learning segmentation and clinical relevance of retropharyngeal edema volume
Objective: Retropharyngeal edema (RPE) on MRI in patients with acute neck infection is associated with disease severity. We explored the potential role of RPE volume as a quantitative marker and developed a convolutional neural network (CNN) for automated RPE volume segmentation. Materials and methods: Volumes of RPE were manually segmented from T2-weighted fat-suppressed Dixon magnetic resonance (MR) images from 244 patients. These volumes were correlated with clinical variables, such as the need for intensive care unit (ICU) admissions, C-reactive protein (CRP) levels, maximal abscess diameter, and length of hospital stay (LOS). Manually segmented masks were used to train a CNN. Results: Patients who required ICU admission had significantly higher RPE volumes than those who did not, and RPE volume outperformed the binary RPE (presence/absence) in classification analysis of ICU admissions. Furthermore, RPE volume correlated positively with LOS, CRP, and maximal abscess diameter. At the slice level, the deep learning (DL)-based model achieved its highest area under the receiver operating characteristic curve (AUROC) in sagittal slices (98.2%) and its highest Dice similarity coefficient in axial slices (0.534). Conclusion: RPE volume is a promising quantitative imaging biomarker associated with relevant clinical outcomes in acute neck infections. Our DL-based model enables automated quantification of RPE volume. Relevance statement: RPE volume provides clinically meaningful information in acute neck infections, outperforming binary classification in predicting disease severity and correlating with key clinical outcomes. Automated DL-based segmentation accurately locates the RPE and provides a moderate quantitative measurement of RPE volume, supporting its potential as a clinical imaging biomarker. Key Points: RPE volume correlated with markers of severe illness and outperformed binary RPE classification. We developed a DL-based algorithm for slice-wise classification and automatic segmentation of RPE. The classification model achieved excellent performance, while segmentation yielded modest Dice similarity coefficients consistent with prior imaging-based tumor segmentation algorithms.
- Open Access
Fast capacity computation for maze-like configurations
We study the conformal capacity ${\rm cap}(Ω,K)$ where $Ω$ is a bounded domain of $\mathbb{R}^2$ and $K$ is a compact connected set in $Ω$. Because the exact numerical value of the capacity is known only in a handful of special cases, it is important to find estimates for the capacity in terms of domain functionals, simpler than the capacity itself. Here, we study condensers of maze-like structure and compute their capacity by means of a high-order $hp$- finite element method. We compare these numerical results to the estimates given by the quasihyperbolic length and perimeter of the compact set. In particular, we consider the behaviour of these value pairs, numerical results and estimates, when the structure parameters vary and the walls of the maze approach the compact set. Over the configurations covered in the numerical experiments, the quasihyperbolic estimates are shown to have the desired asymptotic properties and superior computational efficiencies once the case-specific analysis is completed.Open AccessFast capacity computation for maze-like configurations
We study the conformal capacity ${\rm cap}(Ω,K)$ where $Ω$ is a bounded domain of $\mathbb{R}^2$ and $K$ is a compact connected set in $Ω$. Because the exact numerical value of the capacity is known only in a handful of special cases, it is important to find estimates for the capacity in terms of domain functionals, simpler than the capacity itself. Here, we study condensers of maze-like structure and compute their capacity by means of a high-order $hp$- finite element method. We compare these numerical results to the estimates given by the quasihyperbolic length and perimeter of the compact set. In particular, we consider the behaviour of these value pairs, numerical results and estimates, when the structure parameters vary and the walls of the maze approach the compact set. Over the configurations covered in the numerical experiments, the quasihyperbolic estimates are shown to have the desired asymptotic properties and superior computational efficiencies once the case-specific analysis is completed. - Open Access
- 25 September 2025
- Published by Wiley in Clinical Physiology and Functional Imaging
- Vol. 45 (5) , e70030
Deep learning‐based 3D classification of head and neck cancer PET/MRI: Radiologist comparison and Grad‐CAM interpretability
Purpose: To develop and evaluate a three‐dimensional convolutional neural network for automated classification of PET/MRI images in head and neck cancer (HNC) patients, assessing its performance against radiologist interpretation and its potential as a diagnostic aid.Methods: Data from 202 patients with HNC who underwent 18F‐FDG PET/MRI were used to train and validate PET‐, MRI‐, and PET/MRI‐based models. Of these data, 101 patients were labelled as positive in terms of having HNC, and 101 patients as negative. An additional test set of 20 patients was also evaluated, where 10 patients were labelled as positive and 10 as negative. The model performance was assessed using sensitivity, specificity, accuracy, and AUC. Grad‐CAM was utilised to improve interpretability and classification results on the test set were compared with a radiologist.Results: The PET‐based model achieved an AUC of 0.92 on the test set, with an accuracy of 90%, a sensitivity of 100% and a specificity of 80%. PET/MRI and MRI‐based models underperformed relative to the PET‐based model. The radiologist achieved perfect classification accuracy. Analysis of Grad‐CAM showed that the model classifications are based on real areas of interest. In addition, it gave valuable insight into using similar systems in identifying false positive findings.Conclusion: The PET‐based model demonstrated high sensitivity, indicating its potential as a pre‐screening tool for HNC. However, specificity requires improvement to reduce false‐positive rates. Enhanced datasets and refinement of model architecture will be crucial before clinical adoption. Grad‐CAM provides valuable insights into model decisions, aiding clinical integration.
- 25 September 2025
- Published by Wiley in Clinical Physiology and Functional Imaging
- Vol. 45 (5) , e70030
Open AccessDeep learning‐based 3D classification of head and neck cancer PET/MRI: Radiologist comparison and Grad‐CAM interpretability
Purpose: To develop and evaluate a three‐dimensional convolutional neural network for automated classification of PET/MRI images in head and neck cancer (HNC) patients, assessing its performance against radiologist interpretation and its potential as a diagnostic aid.Methods: Data from 202 patients with HNC who underwent 18F‐FDG PET/MRI were used to train and validate PET‐, MRI‐, and PET/MRI‐based models. Of these data, 101 patients were labelled as positive in terms of having HNC, and 101 patients as negative. An additional test set of 20 patients was also evaluated, where 10 patients were labelled as positive and 10 as negative. The model performance was assessed using sensitivity, specificity, accuracy, and AUC. Grad‐CAM was utilised to improve interpretability and classification results on the test set were compared with a radiologist.Results: The PET‐based model achieved an AUC of 0.92 on the test set, with an accuracy of 90%, a sensitivity of 100% and a specificity of 80%. PET/MRI and MRI‐based models underperformed relative to the PET‐based model. The radiologist achieved perfect classification accuracy. Analysis of Grad‐CAM showed that the model classifications are based on real areas of interest. In addition, it gave valuable insight into using similar systems in identifying false positive findings.Conclusion: The PET‐based model demonstrated high sensitivity, indicating its potential as a pre‐screening tool for HNC. However, specificity requires improvement to reduce false‐positive rates. Enhanced datasets and refinement of model architecture will be crucial before clinical adoption. Grad‐CAM provides valuable insights into model decisions, aiding clinical integration.
- Open Access
- 7 July 2025
- Published by Springer Nature in Network Modeling Analysis in Health Informatics and Bioinformatics
- Vol. 14 (1) , 1-14
Convolutional neural networks for tumor segmentation by cancer type and imaging modality: a systematic review
During the last few decades, a convolutional neural network (CNN) has become a very popular deep learning technique in automatic tumor segmentation of medical images of cancer patients. However, unlike for semantic segmentation of regular photographs, there are few publicly available datasets that can be used to train a CNN to perform tumor segmentation in medical images. Consequently, it is difficult to compare these methods or tell how well they should work for a specific combination of cancer type and imaging modality when trained with a certain amount of data. To answer this question, we analyzed 325 recent articles about CNNs trained for tumor segmentation in order to give a comprehensive overview of the current state of research. The articles study several different types of cancer, including brain tumors and breast, liver, lung, skin, head and neck, prostate, thyroid, cervical, colorectal, pancreatic, kidney, and bladder cancer, imaged with magnetic resonance imaging (MRI), computed tomography, positron emission tomography, ultrasound, and other similar modalities. According to our analysis, the most popular CNN for tumor segmentation is U-Net and its new modifications. Conversely, Mask region-based CNNs are rarely used outside of MRI images. Out of the other CNN designs, SegNet and DeepLabV3 are most common but still significantly less studied than U-Net. Furthermore, several methods have not yet been tested for rarer types of cancer.
- 7 July 2025
- Published by Springer Nature in Network Modeling Analysis in Health Informatics and Bioinformatics
- Vol. 14 (1) , 1-14
Open AccessConvolutional neural networks for tumor segmentation by cancer type and imaging modality: a systematic review
During the last few decades, a convolutional neural network (CNN) has become a very popular deep learning technique in automatic tumor segmentation of medical images of cancer patients. However, unlike for semantic segmentation of regular photographs, there are few publicly available datasets that can be used to train a CNN to perform tumor segmentation in medical images. Consequently, it is difficult to compare these methods or tell how well they should work for a specific combination of cancer type and imaging modality when trained with a certain amount of data. To answer this question, we analyzed 325 recent articles about CNNs trained for tumor segmentation in order to give a comprehensive overview of the current state of research. The articles study several different types of cancer, including brain tumors and breast, liver, lung, skin, head and neck, prostate, thyroid, cervical, colorectal, pancreatic, kidney, and bladder cancer, imaged with magnetic resonance imaging (MRI), computed tomography, positron emission tomography, ultrasound, and other similar modalities. According to our analysis, the most popular CNN for tumor segmentation is U-Net and its new modifications. Conversely, Mask region-based CNNs are rarely used outside of MRI images. Out of the other CNN designs, SegNet and DeepLabV3 are most common but still significantly less studied than U-Net. Furthermore, several methods have not yet been tested for rarer types of cancer.
- Open Access
- 19 June 2025
- Published by Springer Nature in European Radiology Experimental
- Vol. 9 (1) , 1-11
Deep learning detects retropharyngeal edema on MRI in patients with acute neck infections
Background: In acute neck infections, magnetic resonance imaging (MRI) shows retropharyngeal edema (RPE), which is a prognostic imaging biomarker for a severe course of illness. This study aimed to develop a deep learning-based algorithm for the automated detection of RPE. Methods: We developed a deep neural network consisting of two parts using axial T2-weighted water-only Dixon MRI images from 479 patients with acute neck infections annotated by radiologists at both slice and patient levels. First, a convolutional neural network (CNN) classified individual slices; second, an algorithm classified patients based on a stack of slices. Model performance was compared with the radiologists’ assessment as a reference standard. Accuracy, sensitivity, specificity, and area under receiver operating characteristic curve (AUROC) were calculated. The proposed CNN was compared with InceptionV3, and the patient-level classification algorithm was compared with traditional machine learning models. Results: Of the 479 patients, 244 (51%) were positive and 235 (49%) negative for RPE. Our model achieved accuracy, sensitivity, specificity, and AUROC of 94.6%, 83.3%, 96.2%, and 94.1% at the slice level, and 87.4%, 86.5%, 88.2%, and 94.8% at the patient level, respectively. The proposed CNN was faster than InceptionV3 but equally accurate. Our patient classification algorithm outperformed traditional machine learning models. Conclusion: A deep learning model, based on weakly annotated data and computationally manageable training, achieved high accuracy for automatically detecting RPE on MRI in patients with acute neck infections. Relevance statement: Our automated method for detecting relevant MRI findings was efficiently trained and might be easily deployed in practice to study clinical applicability. This approach might improve early detection of patients at high risk for a severe course of acute neck infections. Key Points: Deep learning automatically detected retropharyngeal edema on MRI in acute neck infections. Areas under the receiver operating characteristic curve were 94.1% at the slice level and 94.8% at the patient level. The proposed convolutional neural network was lightweight and required only weakly annotated data. Graphical
- 19 June 2025
- Published by Springer Nature in European Radiology Experimental
- Vol. 9 (1) , 1-11
Open AccessDeep learning detects retropharyngeal edema on MRI in patients with acute neck infections
Background: In acute neck infections, magnetic resonance imaging (MRI) shows retropharyngeal edema (RPE), which is a prognostic imaging biomarker for a severe course of illness. This study aimed to develop a deep learning-based algorithm for the automated detection of RPE. Methods: We developed a deep neural network consisting of two parts using axial T2-weighted water-only Dixon MRI images from 479 patients with acute neck infections annotated by radiologists at both slice and patient levels. First, a convolutional neural network (CNN) classified individual slices; second, an algorithm classified patients based on a stack of slices. Model performance was compared with the radiologists’ assessment as a reference standard. Accuracy, sensitivity, specificity, and area under receiver operating characteristic curve (AUROC) were calculated. The proposed CNN was compared with InceptionV3, and the patient-level classification algorithm was compared with traditional machine learning models. Results: Of the 479 patients, 244 (51%) were positive and 235 (49%) negative for RPE. Our model achieved accuracy, sensitivity, specificity, and AUROC of 94.6%, 83.3%, 96.2%, and 94.1% at the slice level, and 87.4%, 86.5%, 88.2%, and 94.8% at the patient level, respectively. The proposed CNN was faster than InceptionV3 but equally accurate. Our patient classification algorithm outperformed traditional machine learning models. Conclusion: A deep learning model, based on weakly annotated data and computationally manageable training, achieved high accuracy for automatically detecting RPE on MRI in patients with acute neck infections. Relevance statement: Our automated method for detecting relevant MRI findings was efficiently trained and might be easily deployed in practice to study clinical applicability. This approach might improve early detection of patients at high risk for a severe course of acute neck infections. Key Points: Deep learning automatically detected retropharyngeal edema on MRI in acute neck infections. Areas under the receiver operating characteristic curve were 94.1% at the slice level and 94.8% at the patient level. The proposed convolutional neural network was lightweight and required only weakly annotated data. Graphical
- Open Access
- 31 May 2025
- Published by Springer Nature in Network Modeling Analysis in Health Informatics and Bioinformatics
- Vol. 14 (1) , 1-16
New 3D segmentation algorithm for modeling of kidney in positron emission tomography images
Background: $$^{15}$$ O-water positron emission tomography (PET) imaging enables noninvasive quantification of renal blood flow. While there are several existing methods automatically locating the volume of interest (VOI) for a kidney, separate cortex and medulla VOIs are needed for PET modeling due their functional differences. To assist in this, we introduce a new three-dimensional segmentation algorithm for kidney. Materials and methods: Our algorithm chooses an initial kidney VOI, finds its longitudinal axis and the direction of the non-convex part, and then creates cortex and medulla VOIs and removes the renal vessel area from them. We evaluated the algorithm by using it to define cortex and medulla VOIs of left and right kidneys in dynamic total-body $$^{15}$$ O-water PET images of 35 human patients. For all the 70 kidneys, we plotted the cross-section of the VOIs from three different anatomic directions and asked two expert physicians to assess their quality. Additionally, we computed cortical and medullary renal blood flow estimates by fitting a compartment model to the mean time-activity curves of our VOIs. Results: According to the evaluation by the physicians, the cortex and medulla VOIs were mostly correct in all three directions for 78.6% of the total 70 kidneys and correct in at least one direction for 94.3% of the kidneys. The segmentation inaccuracies were typically caused by the algorithm placing cortex VOI partially outside of the kidney or in the medulla. However, regardless of these inaccuracies, all the VOIs were accurate enough to be used for compartment modeling. The resulting cortical and medullary blood flow were very close to the values reported in earlier studies with similar patient populations. Conclusion: Our proposed algorithm can be used to create an automatic pipeline for accurate quantification of cortical and medullary tracer perfusion after a PET scan.- 31 May 2025
- Published by Springer Nature in Network Modeling Analysis in Health Informatics and Bioinformatics
- Vol. 14 (1) , 1-16
Open AccessNew 3D segmentation algorithm for modeling of kidney in positron emission tomography images
Background: $$^{15}$$ O-water positron emission tomography (PET) imaging enables noninvasive quantification of renal blood flow. While there are several existing methods automatically locating the volume of interest (VOI) for a kidney, separate cortex and medulla VOIs are needed for PET modeling due their functional differences. To assist in this, we introduce a new three-dimensional segmentation algorithm for kidney. Materials and methods: Our algorithm chooses an initial kidney VOI, finds its longitudinal axis and the direction of the non-convex part, and then creates cortex and medulla VOIs and removes the renal vessel area from them. We evaluated the algorithm by using it to define cortex and medulla VOIs of left and right kidneys in dynamic total-body $$^{15}$$ O-water PET images of 35 human patients. For all the 70 kidneys, we plotted the cross-section of the VOIs from three different anatomic directions and asked two expert physicians to assess their quality. Additionally, we computed cortical and medullary renal blood flow estimates by fitting a compartment model to the mean time-activity curves of our VOIs. Results: According to the evaluation by the physicians, the cortex and medulla VOIs were mostly correct in all three directions for 78.6% of the total 70 kidneys and correct in at least one direction for 94.3% of the kidneys. The segmentation inaccuracies were typically caused by the algorithm placing cortex VOI partially outside of the kidney or in the medulla. However, regardless of these inaccuracies, all the VOIs were accurate enough to be used for compartment modeling. The resulting cortical and medullary blood flow were very close to the values reported in earlier studies with similar patient populations. Conclusion: Our proposed algorithm can be used to create an automatic pipeline for accurate quantification of cortical and medullary tracer perfusion after a PET scan.