Miika Tapio Nieminen
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Miika Tapio Nieminen has a total of 293 co-authors. Here are the top co-authors by number of shared publications:
- Simo Saarakkala
Research Unit of Health Sciences and Technology, Medical Research Center, Oulu University Hospital and University of Oulu, Oulu, Finland
17 shared publications - Mikko Johannes Nissi
Department of Technical Physics University of Eastern Finland Kuopio Finland
12 shared publications - Victor Casula
Research Unit of Health Sciences and Technology, Faculty of Medicine University of Oulu Oulu Finland
12 shared publications - Mikael Asko Kaarlo Juntunen
Research Unit of Health Sciences and Technology, University of Oulu, Oulu, Finland
9 shared publications - Satu Irene Inkinen
University of Oulu
9 shared publications - Ari Heinonen
Faculty of Sport and Health Sciences, University of Jyvaskyla, Jyvaskyla, Finland
8 shared publications
Latest Publications
- 25 September 2026
- Published by Elsevier in Radiography
- Vol. 32 (7)
Reply to the Letter to the Editor: “From visualization to clinical practice: Considerations on the educational effectiveness of an X-ray scatter mobile application”
- 25 September 2026
- Published by Elsevier in Radiography
- Vol. 32 (7)
Reply to the Letter to the Editor: “From visualization to clinical practice: Considerations on the educational effectiveness of an X-ray scatter mobile application”
- Open Access
- 17 June 2026
- Published by Springer Nature
- p. 644-647
AI-Assisted Documentation in Social Welfare Services: Insights from a Pilot Conducted in the Wellbeing Services County of North Ostrobothnia
Artificial intelligence (AI) and large language models (LLMs) are increasingly explored as tools to reduce administrative burden in health and social care. In Finland, wellbeing services counties carry extensive responsibilities for organizing social welfare and healthcare services, including documentation practices that are often time-consuming, complex, and essential for service quality. While LLM applications have progressed rapidly in healthcare settings, experimentation within social welfare work remains limited. This paper reports findings from a pilot conducted in the wellbeing services county of North Ostrobothnia, where a generative AI–based documentation tool was tested by social welfare professionals. The results indicate improvements in documentation efficiency, uniformity, and interaction quality, but also highlight ethical, legal, and sociotechnical challenges concerning data protection, transparency, professional accountability, and contextual applicability. The pilot suggests that LLMs can support social welfare practice when implemented responsibly, with strong professional oversight and attention to contextual factors.
- 17 June 2026
- Published by Springer Nature
- p. 644-647
Open AccessAI-Assisted Documentation in Social Welfare Services: Insights from a Pilot Conducted in the Wellbeing Services County of North Ostrobothnia
Artificial intelligence (AI) and large language models (LLMs) are increasingly explored as tools to reduce administrative burden in health and social care. In Finland, wellbeing services counties carry extensive responsibilities for organizing social welfare and healthcare services, including documentation practices that are often time-consuming, complex, and essential for service quality. While LLM applications have progressed rapidly in healthcare settings, experimentation within social welfare work remains limited. This paper reports findings from a pilot conducted in the wellbeing services county of North Ostrobothnia, where a generative AI–based documentation tool was tested by social welfare professionals. The results indicate improvements in documentation efficiency, uniformity, and interaction quality, but also highlight ethical, legal, and sociotechnical challenges concerning data protection, transparency, professional accountability, and contextual applicability. The pilot suggests that LLMs can support social welfare practice when implemented responsibly, with strong professional oversight and attention to contextual factors.
- Open Access
- 17 June 2026
- Published by Springer Nature
- p. 172-176
Establishing an AI Competence Center in a Public Healthcare Organization: A Multidisciplinary Learning Pathway
The rapid evolution of Artificial Intelligence (AI) requires public healthcare organizations to balance innovation with strict regulatory compliance. This paper describes the establishment of an AI competence center at the Wellbeing Services County of North Ostrobothnia (Pohde) and at Oulu University Hospital (OYS). The AI competence center was designed to bridge the gap between wellbeing services county organization’s needs and technological capabilities through a multidisciplinary structure involving the University of Oulu and industry partners. In addition to establishment activities, we present the AI competence center’s “learning pathway,” a four-phase methodology consisting of service delivery domain level scouting, market alignment, feasibility ranking, and hackathon-based validation. The results indicate that a centralized, cross-functional competence center accelerates the identification and operationalization of high-value AI solutions.
- 17 June 2026
- Published by Springer Nature
- p. 172-176
Open AccessEstablishing an AI Competence Center in a Public Healthcare Organization: A Multidisciplinary Learning Pathway
The rapid evolution of Artificial Intelligence (AI) requires public healthcare organizations to balance innovation with strict regulatory compliance. This paper describes the establishment of an AI competence center at the Wellbeing Services County of North Ostrobothnia (Pohde) and at Oulu University Hospital (OYS). The AI competence center was designed to bridge the gap between wellbeing services county organization’s needs and technological capabilities through a multidisciplinary structure involving the University of Oulu and industry partners. In addition to establishment activities, we present the AI competence center’s “learning pathway,” a four-phase methodology consisting of service delivery domain level scouting, market alignment, feasibility ranking, and hackathon-based validation. The results indicate that a centralized, cross-functional competence center accelerates the identification and operationalization of high-value AI solutions.
- Open Access
- 17 June 2026
- Published by Springer Nature
- p. 19-30
Computer Vision-Based Breast Volume Quantification Using Digital Mammography Data with Manual Measurement Validation
Mammographic images are commonly used in cancer screening protocols, yet there is a lack of automated tools evaluating breast volume from the images, which could help preoperative planning of breast surgery or epidemiological research. The study aimed at creating and validating an automated workflow utilizing computer vision for quantifying breast volume from a large dataset of digital mammographic images. The study was conducted in two phases. First, an automated workflow for analyzing breast volume was developed using Python-based computer vision libraries. This was followed by the validation and analysis of the method by utilizing seventy-three mammographic images of women undergoing preoperative mammography. The performance of the automated volume calculation method was evaluated and validated with manually measured breast dimensions. In a retrospective register study, a total of seventy-three mammograms from 41 females ages 35–82 years (mean 60 years) were analyzed showing effective analysis process and excellent accuracy of the automated computer vision approach in estimating breast volume from mammographic data. The mean difference (bias) between manually measured and computer vision approach volumes was minimal (0.01 cm 3 ) resulting in an excellent correlation (R = 1.0, p < 0.001) between the techniques. The proposed automated method provides an accurate, fast, and cost-effective solution for obtaining breast volume measurements from large-scale mammography datasets, and it could be useful for preoperative planning and research on breast health.
- 17 June 2026
- Published by Springer Nature
- p. 19-30
Open AccessComputer Vision-Based Breast Volume Quantification Using Digital Mammography Data with Manual Measurement Validation
Mammographic images are commonly used in cancer screening protocols, yet there is a lack of automated tools evaluating breast volume from the images, which could help preoperative planning of breast surgery or epidemiological research. The study aimed at creating and validating an automated workflow utilizing computer vision for quantifying breast volume from a large dataset of digital mammographic images. The study was conducted in two phases. First, an automated workflow for analyzing breast volume was developed using Python-based computer vision libraries. This was followed by the validation and analysis of the method by utilizing seventy-three mammographic images of women undergoing preoperative mammography. The performance of the automated volume calculation method was evaluated and validated with manually measured breast dimensions. In a retrospective register study, a total of seventy-three mammograms from 41 females ages 35–82 years (mean 60 years) were analyzed showing effective analysis process and excellent accuracy of the automated computer vision approach in estimating breast volume from mammographic data. The mean difference (bias) between manually measured and computer vision approach volumes was minimal (0.01 cm 3 ) resulting in an excellent correlation (R = 1.0, p < 0.001) between the techniques. The proposed automated method provides an accurate, fast, and cost-effective solution for obtaining breast volume measurements from large-scale mammography datasets, and it could be useful for preoperative planning and research on breast health.
- Open Access
- 17 June 2026
- Published by Springer Nature
- p. 3-18
A Multi-filter and Multi-scale U-Net for Cone-Beam Computed Tomography
Efficient 3D cone-beam computed tomography (CBCT) reconstructions typically demand substantial hardware capabilities for both training and execution. This paper introduces a multi-filter and multi-scale U-Net (mU-Net) architecture designed for post-processing CBCT images by incorporating multiple input channels, each processed with distinct frequency filters or reconstruction resolutions. We present a detailed comparison between the mU-Net and a baseline U-Net for low-dose and clinical dose denoising scenarios, reporting key metrics such as Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and inference time. The mU-Net showed only marginal quantitative differences compared with the baseline U-Net: it performed slightly better in the clinical dose scenario, while the baseline U-Net performed better in the low-dose scenario. The mU-Net required approximately 13% more time per epoch during training and 13.08 s per volume for FDK reconstruction and inference, compared with 10.43 s per volume for the baseline U-Net. In the clinical dose scenario, the mU-Net showed slightly improved visual noise texture and artifact suppression, but these observations were not supported by dedicated texture or observer-study metrics. Notably, the mU-Net and baseline U-Net approaches placed 3rd in the ICASSP 2024 3D-CBCT grand challenge, demonstrating that a properly trained U-Net remains competitive. This study shows that multi-filter and multi-scale inputs can be used for CBCT post-processing under limited computational resources, but the gains over a baseline U-Net were modest.
- 17 June 2026
- Published by Springer Nature
- p. 3-18
Open AccessA Multi-filter and Multi-scale U-Net for Cone-Beam Computed Tomography
Efficient 3D cone-beam computed tomography (CBCT) reconstructions typically demand substantial hardware capabilities for both training and execution. This paper introduces a multi-filter and multi-scale U-Net (mU-Net) architecture designed for post-processing CBCT images by incorporating multiple input channels, each processed with distinct frequency filters or reconstruction resolutions. We present a detailed comparison between the mU-Net and a baseline U-Net for low-dose and clinical dose denoising scenarios, reporting key metrics such as Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and inference time. The mU-Net showed only marginal quantitative differences compared with the baseline U-Net: it performed slightly better in the clinical dose scenario, while the baseline U-Net performed better in the low-dose scenario. The mU-Net required approximately 13% more time per epoch during training and 13.08 s per volume for FDK reconstruction and inference, compared with 10.43 s per volume for the baseline U-Net. In the clinical dose scenario, the mU-Net showed slightly improved visual noise texture and artifact suppression, but these observations were not supported by dedicated texture or observer-study metrics. Notably, the mU-Net and baseline U-Net approaches placed 3rd in the ICASSP 2024 3D-CBCT grand challenge, demonstrating that a properly trained U-Net remains competitive. This study shows that multi-filter and multi-scale inputs can be used for CBCT post-processing under limited computational resources, but the gains over a baseline U-Net were modest.
- Open Access
- 17 June 2026
- Published by Springer Nature
- p. 538-556
Feasibility of Deep Learning-Based Classification of Lumbar T$_2$-Weighted MRIs in Subjects With and Without Low Back Pain SymptomsBackground: Low back pain (LBP) is a prevalent condition. Magnetic resonance imaging (MRI) enables visualization of lumbar spinal structures, but the relationship between MRI findings and symptoms is complex, as degenerative changes are also frequently observed in asymptomatic individuals. Purpose: This study investigates the feasibility of deep learning for automated classification of lumbar spine MRIs into symptomatic and asymptomatic cases. The aim was to assess whether symptom-related information can be learned from T $$_2$$ -weighted MRI data. Study Design: A deep learning-based classification approach using a pre-trained ResNet50 convolutional neural network (CNN) was developed to distinguish between symptomatic and asymptomatic cases. Methods: Sagittal T $$_2$$ -weighted MRIs from the Northern Finland Birth Cohort 1966 dataset were used to train, validate, and test the ResNet50-based model for LBP classification. The impact of varying the number of sagittal slices per subject was also evaluated. Results: The best classification performance was achieved using a modified pre-trained ResNet50 network. The model attained a Balanced Accuracy of $$0.713 \pm 0.014$$ , an Average Precision of $$0.494 \pm 0.013$$ , and an Area Under the Receiver Operating Characteristic Curve (AUC-ROC) of $$0.724 \pm 0.004$$ . Binary classification was more effective in predicting asymptomatic cases, while incorporating Pfirrmann grading improved symptomatic case predictions. Conclusions: Deep learning-based analysis of lumbar spine MRI showed modest ability to distinguish symptomatic and asymptomatic cases in this cohort. The results support the feasibility of this exploratory approach while highlighting important clinical limitations.- 17 June 2026
- Published by Springer Nature
- p. 538-556
Open AccessFeasibility of Deep Learning-Based Classification of Lumbar T$_2$-Weighted MRIs in Subjects With and Without Low Back Pain SymptomsBackground: Low back pain (LBP) is a prevalent condition. Magnetic resonance imaging (MRI) enables visualization of lumbar spinal structures, but the relationship between MRI findings and symptoms is complex, as degenerative changes are also frequently observed in asymptomatic individuals. Purpose: This study investigates the feasibility of deep learning for automated classification of lumbar spine MRIs into symptomatic and asymptomatic cases. The aim was to assess whether symptom-related information can be learned from T $$_2$$ -weighted MRI data. Study Design: A deep learning-based classification approach using a pre-trained ResNet50 convolutional neural network (CNN) was developed to distinguish between symptomatic and asymptomatic cases. Methods: Sagittal T $$_2$$ -weighted MRIs from the Northern Finland Birth Cohort 1966 dataset were used to train, validate, and test the ResNet50-based model for LBP classification. The impact of varying the number of sagittal slices per subject was also evaluated. Results: The best classification performance was achieved using a modified pre-trained ResNet50 network. The model attained a Balanced Accuracy of $$0.713 \pm 0.014$$ , an Average Precision of $$0.494 \pm 0.013$$ , and an Area Under the Receiver Operating Characteristic Curve (AUC-ROC) of $$0.724 \pm 0.004$$ . Binary classification was more effective in predicting asymptomatic cases, while incorporating Pfirrmann grading improved symptomatic case predictions. Conclusions: Deep learning-based analysis of lumbar spine MRI showed modest ability to distinguish symptomatic and asymptomatic cases in this cohort. The results support the feasibility of this exploratory approach while highlighting important clinical limitations. - Open Access
- 13 June 2026
- Published by Elsevier in Osteoarthritis and Cartilage Open
- Vol. 8 (3)
Associations of hemoglobin levels with structural knee MRI findings at 33 years of age in a general population-based birth cohort
- 13 June 2026
- Published by Elsevier in Osteoarthritis and Cartilage Open
- Vol. 8 (3)
Open AccessAssociations of hemoglobin levels with structural knee MRI findings at 33 years of age in a general population-based birth cohort
- Open Access
- 11 June 2026
- Published by Elsevier in Radiography
- Vol. 32 (5) , 103468
Impact of an X-ray scatter mobile application on radiation protection training: A user experience study
- 11 June 2026
- Published by Elsevier in Radiography
- Vol. 32 (5) , 103468
Open AccessImpact of an X-ray scatter mobile application on radiation protection training: A user experience study
- Open Access
- 25 March 2026
- Published by Wiley in Journal of Magnetic Resonance Imaging
- Vol. 64 (1) , 90-109
Clinical Feasibility of Deep Learning Contrast Synthesis From MR Fingerprinting in Knee Osteoarthritis
Background: Magnetic Resonance Fingerprinting (MRF) enables rapid quantitative parameter mapping from which synthetic clinical contrast images can be derived using deep learning (DL). Purpose: This study evaluates the reliability and interchangeability of MRF‐derived synthetic knee MRI relative to conventional MRI in patients with osteoarthritis. Study Type: Prospective single‐center comparative study. Subjects: Between March 2022 and 2023, 78 participants (54 females, mean age 57.2, range 33–78 years) with knee osteoarthritis. Field Strength/Sequence: 3.0 T; proton density weighted (PDw) imaging, T2‐weighted fat‐saturated (T2w fs) imaging, and MRF. Assessment: U‐Nets were trained to produce synthetic contrasts from MRF data. Three musculoskeletal radiologists performed MRI OsteoArthritis Knee Score (MOAKS) assessments and image quality evaluation using a Likert scale (1–5). Statistical Tests: The inter‐rater and inter‐method reliability were evaluated using prevalence‐and‐bias‐adjusted kappa (PABAK), and percentages of exact matches. Image quality scores were compared using the Wilcoxon test. The limit of statistical significance was set at p < 0.05 and no multiple comparisons corrections were applied. Results: The inter‐rater reliability for synthetic MR images varied between 0.980–0.994 (CI 0.975–0.997; exact matches 77.%–89.2%) and for conventional MR images between 0.979 and 0.994 (CI 0.973–0.997; exact matches 75.3%–89.5%). Inter‐method reliability between synthetic and conventional MR images was near‐perfect: mean PABAK‐values and exact matches were 0.927 and 77.9% for cartilage, 0.915 and 89.7% for bone marrow lesions, 0.922 and 65.2% for osteophytes, 0.950 and 72.9% for meniscus pathology, 0.934 and 66.6% for effusion, and 0.857 and 92.5% for Baker's cyst. Average Likert scores were significantly better for conventional than synthetic images: 4.5 vs. 3.9 for PDw and 4.1 vs. 3.2 for T2w fs. Data Conclusion: MRF‐derived DL‐based synthetic clinical contrasts provide excellent inter‐rater reliability and interchangeability against conventional MR sequences in knee OA; however, image quality needs further development. Evidence Level: 2. Stage of Technical Efficacy: 2.
- 25 March 2026
- Published by Wiley in Journal of Magnetic Resonance Imaging
- Vol. 64 (1) , 90-109
Open AccessClinical Feasibility of Deep Learning Contrast Synthesis From MR Fingerprinting in Knee Osteoarthritis
Background: Magnetic Resonance Fingerprinting (MRF) enables rapid quantitative parameter mapping from which synthetic clinical contrast images can be derived using deep learning (DL). Purpose: This study evaluates the reliability and interchangeability of MRF‐derived synthetic knee MRI relative to conventional MRI in patients with osteoarthritis. Study Type: Prospective single‐center comparative study. Subjects: Between March 2022 and 2023, 78 participants (54 females, mean age 57.2, range 33–78 years) with knee osteoarthritis. Field Strength/Sequence: 3.0 T; proton density weighted (PDw) imaging, T2‐weighted fat‐saturated (T2w fs) imaging, and MRF. Assessment: U‐Nets were trained to produce synthetic contrasts from MRF data. Three musculoskeletal radiologists performed MRI OsteoArthritis Knee Score (MOAKS) assessments and image quality evaluation using a Likert scale (1–5). Statistical Tests: The inter‐rater and inter‐method reliability were evaluated using prevalence‐and‐bias‐adjusted kappa (PABAK), and percentages of exact matches. Image quality scores were compared using the Wilcoxon test. The limit of statistical significance was set at p < 0.05 and no multiple comparisons corrections were applied. Results: The inter‐rater reliability for synthetic MR images varied between 0.980–0.994 (CI 0.975–0.997; exact matches 77.%–89.2%) and for conventional MR images between 0.979 and 0.994 (CI 0.973–0.997; exact matches 75.3%–89.5%). Inter‐method reliability between synthetic and conventional MR images was near‐perfect: mean PABAK‐values and exact matches were 0.927 and 77.9% for cartilage, 0.915 and 89.7% for bone marrow lesions, 0.922 and 65.2% for osteophytes, 0.950 and 72.9% for meniscus pathology, 0.934 and 66.6% for effusion, and 0.857 and 92.5% for Baker's cyst. Average Likert scores were significantly better for conventional than synthetic images: 4.5 vs. 3.9 for PDw and 4.1 vs. 3.2 for T2w fs. Data Conclusion: MRF‐derived DL‐based synthetic clinical contrasts provide excellent inter‐rater reliability and interchangeability against conventional MR sequences in knee OA; however, image quality needs further development. Evidence Level: 2. Stage of Technical Efficacy: 2.
- Open Access
- 14 March 2026
- Published by Elsevier in Osteoarthritis and Cartilage Open
- Vol. 8 (2) , 100780
Associations of health parameters at 16 years of age with structural knee MRI findings at 33 years of age in a general population-based birth cohort
- 14 March 2026
- Published by Elsevier in Osteoarthritis and Cartilage Open
- Vol. 8 (2) , 100780
Open AccessAssociations of health parameters at 16 years of age with structural knee MRI findings at 33 years of age in a general population-based birth cohort