Laurentius Oscar Osapoetra

Introduction: Glioblastoma (GBM) commonly recurs close to the original gross tumor volume (GTV), suggesting treatment-resistant subregions may be present before radiologically visible progression. This study investigated whether deep learning on apparent diffusion coefficient (ADC) maps could generate pixel-wise probability maps of treatment-resistant regions within the GTV. Methods: The training/validation set included images of 64 in-house patients after radiation fraction 10 (Fx10) and 40 external patients before treatment used as an auxiliary cross-domain training cohort; the held-out test set included 16 in-house Fx10 patients. Resistant regions were defined as the overlap between the GTV from scans acquired at radiation fraction 10 and subsequent recurrence after finishing treatment. Models were trained on 2D GTV-cropped slices with patient-level separation to prevent data leakage. Single- and multimodal MRI inputs were compared, and ADC-based models (including baseline UNet, pretrained ResNet34, BiomedCLIP-assisted UNet, and a BLIP image-text UNet) were later evaluated. Performance was evaluated using conventional metrics, including ROC AUC, PR AUC, F1-score, recall, specificity, and precision. Results: ADC was the most informative single modality, achieving 69% and 80% for Receiver Operating Characteristic (ROC)- and Precision-Recall (PR)-Area Under the Curve (AUC), respectively. In the ADC-based architecture comparison, no model was consistently superior across all metrics. The baseline UNet achieved the highest accuracy, F1-score, and recall, whereas pretrained ResNet34 and BLIP-assisted models showed slightly higher ROC AUC than the baseline UNet (both 70%). Error analysis showed that performance decreased substantially for small or sparse resistant regions, and false-negative errors were enriched near resistant-region and GTV boundaries. Conclusions: ADC-based deep learning showed feasibility for pixel-wise spatial mapping of treatment-resistant glioblastoma regions within the GTV. A UNet trained from scratch provided the strongest sensitivity/F1 trade-off, whereas pretrained and VLM-assisted models provided only metric-specific benefits. Model comparisons showed that BiomedCLIP-assisted and BLIP image-text UNet models did not clearly improve fine spatial segmentation performance over the baseline UNet.

Background: A Phase I trial to study the efficacy and safety of MR-guided FUS-MB (focused ultrasound–microbubble) treatment combined with radiotherapy (RT) using a fully electronically MRI-guided steerable focused ultrasound system was performed. The investigation here carried out an exploratory quantitative analysis of images for the first ten participants who were treated with MRgFUS-MB therapy, identifying texture-based MRI features that changed significantly after treatment and indicating probable treatment-related changes in tumour morphology. Methods and Findings: Ten patients with stage I-IV breast cancer (Age: 64 ± 10.8) underwent 2 FUS-MB therapies throughout their radiation treatments, as deemed appropriate by a multidisciplinary team. MRI scans were performed prior to treatment and at 1- and 3-month follow-up. The tumours were segmented, and 2D and 3D features, including shape, first-order and texture features, were extracted for original and wavelet-filtered T1-weighted fat-saturated MR images. The overall significance of radiomic features in association with morphological changes in the target tumour over 1-month and 3-month follow-up in comparison to the first treatment day was investigated. Results: The five most significant features (p < 0.05) were selected using a sample-related t-test, out of which most features were wavelet-filtered first-order features. First-order wavelet-filtered robust mean absolute deviation (RMAD) emerged as the most consistent significant feature for both 2D and 3D features with sampled grid spacings of 0.5 mm and 1 mm. The volume of the tumour demonstrated a decline for 80% of patients over 3 months, as measured by the 3D mesh volume of the tumour ROI. There was a significant decline in tumour size and image heterogeneity, indicating plausible changes in the tumour microenvironment possibly related to the treatment. Conclusions: Previous preclinical studies and clinical evaluations have demonstrated the efficacy of MR-guided FUS-MB therapy. MRI image-based features from the T1-weighted MR scans over various follow-up times provide potential evidence of changes in tumour tissue heterogeneity that might be treatment-related. MRgFUS MB therapy can possibly evolve as a mode of radiotherapy enhancement in the future, which warrants further controlled validation. The availability of adequate data in the future might possibly lead to the identification of image biomarkers related to treatment response, thus providing better clinical evaluations.

Background/Objectives: We aimed to evaluate the generalizability of QUS spectral parametric imaging radiomics for breast tumor characterization in a large cohort through methodological optimization and rigorous validation, including comprehensive feature extraction and evaluation of multiple classifiers, addressing limitations of previous studies potentially affected by data leakage. Methods: This study included 277 participants (185 malignant cases (median age, 51 years [IQR: 44–63 years]) and 92 benign cases (median age, 46 years [IQR: 38–51 years]) with breast masses, acquired between September 2014 and October 2021. QUS spectroscopic analysis resulted in five maps, from which first-order statistical, various textural, and morphological features were extracted from both the tumor core and a surrounding 5 mm margin. The ground truth label was determined from histopathological analysis. Predictive models were developed to distinguish malignant from benign lesions. Their generalization performance was assessed using hold-out validation. Beyond assessing models’ performance, SHapley Additive exPlanations (SHAP) analysis identified the most influential features to the predictions. Results: 329 radiomics features demonstrated statistically significant differences (p-values < 0.00005). Averaged across 50 partitions, SVM-Linear models produced test performance of 82 ± 8% (CI: 59–97) recall, 79 ± 9% (CI: 50–100) specificity, and 0.87 ± 0.05 (CI: 0.71–0.98) area under the receiver operating characteristic curve (AUROC). The best single partition performance with an SVM-Linear model was 92% recall, 89% specificity, and 0.97 AUROC. Conclusions: This work establishes a performance benchmark for handcrafted radiomic features derived from QUS spectral parametric images in breast lesion characterization, demonstrating its strong generalization and servicing as a reference for future model development.

This study assessed the level of agreement between quantitative ultrasound (QUS) feature estimates derived from ultrasound images of breast tumors in women with locally advanced breast cancer (LABC) produced using a cart-based and a handheld ultrasound system. Thirty LABC patients receiving neoadjuvant chemotherapy were imaged at two separate times: a pre-treatment ‘baseline’ time point, and four weeks after the start of chemotherapy. Three sets of QUS features were produced using the reference phantom technique, one for each imaging time and a third set calculated by taking the differences in feature estimates between times. Cross-system statistical testing using the Wilcoxon signed-rank test was performed for each feature set to assess the level of feature estimate agreement between ultrasound systems. The Bland–Altman method was employed to graphically assess feature sets for systematic skew. The range of p-values was 4.50 × 10⁻¹¹ to 0.277 for the baseline features, 2.77 × 10⁻⁵ to 0.865 for the week 4 features, and 2.03 × 10⁻⁹ to 1 for the feature differences. For the feature differences, all five of the primary QUS features (MBF, SS, SI, ASD, AAC) were found to be in agreement between the two scanner types at the 5% confidence level. For the baseline feature set and week 4 feature set, 0 out of 5 and 3 out of 5 of the primary features were found to be in agreement, respectively. Of the 20 QUS texture features examined, the number and proportion of the total for each feature set which were found to have statistically significant similarity in their sample medians at the 5% confidence level were as follows: 2 out of 20 (10%) for the baseline features; 17 out of 20 (85%) for the week 4 features; and 12 out of 20 (60%) for the feature differences. The specific texture features found to be in agreement varied between QUS-specific feature sets. Overall, a moderate level of agreement between sets of feature differences produced using the two systems was demonstrated.

Background/Objectives: To evaluate the ability of radiomics analysis of QUS spectral parametric imaging to non-invasively differentiate intermediate-to-high-risk from low-risk Oncotype DX^TM Recurrence Score (ODXRS). Methods: This prospective study included 31 participants (21 intermediate-to-high-risk ODXRS (median age, 56 years [IQR: 49–68 years]) and 10 low-risk ODXRS (median age, 52 years [IQR: 48–58 years])) presenting with ER+ HER2− invasive breast masses acquired between September 2015 and August 2024. Quantitative ultrasound (QUS) spectroscopy produced five spectral maps, from which radiomics features (including statistical, texture, and morphological measures) were extracted from the tumor core and a 5 mm margin. The ground truth label was determined from thresholding the ODXRS. A multivariate predictive model was developed to differentiate intermediate-to-high-risk ODXRS from low-risk ODXRS, with performance assessed via nested leave-one-out cross-validation (LOOCV). Results: A nested leave-one-out cross-validation (LOOCV) analysis demonstrated the generalization performance of a four-feature model. The support vector machine (SVM-RBF) classifier achieved 86% recall, 100% specificity, 93% balanced accuracy, and an area under the receiver operating characteristic curve (AUROC) of 0.95 (CI = 0.88–1.00) in identifying intermediate-to-high-risk versus low-risk ODXRS. Conclusions: The preliminary results suggest the potential radiomics-based model of ODXRS in predicting the risks of recurrence. The results warrant further investigation on a larger cohort. This framework can be a useful surrogate for participants for whom ODX testing is neither affordable nor available.

Background : Response to neoadjuvant chemotherapy (NAC) is a key prognostic indicator in breast cancer, yet current assessment relies on postoperative pathology. This study investigated the use of deep features derived from pre-treatment MRI and CT scans, in conjunction with clinical variables, to predict treatment response a priori. Methods: Two response endpoints were analyzed: pathologic complete response (pCR) versus non-pCR, and responders versus non-responders, with response defined as a reduction in tumor size of at least 30%. Intratumoral and peritumoral segmentations were generated on contrast-enhanced T1-weighted (CE-T1) and T2-weighted MRI, as well as contrast-enhanced CT images of tumors. Deep features were extracted from these regions using ResNet10, ResNet18, ResNet34, and ResNet50 architectures pre-trained with MedicalNet. Handcrafted radiomic features were also extracted for comparison. Feature selection was conducted with minimum redundancy maximum relevance (mRMR) followed by recursive feature elimination (RFE), and classification was performed using XGBoost across ten independent data partitions. Results: A total of 177 patients were analyzed in this study. ResNet34-derived features achieved the highest overall classification performance under both criteria, outperforming handcrafted features and deep features from other ResNet architectures. For distinguishing pCR from non-pCR, ResNet34 achieved a balanced accuracy of 81.6%, whereas handcrafted radiomics achieved 77.9%. For distinguishing responders from non-responders, ResNet34 achieved a balanced accuracy of 73.5%, compared with 70.2% for handcrafted radiomics. Conclusions: Deep features extracted from routinely acquired MRI and CT, when combined with clinical information, improve the prediction of NAC response in breast cancer. This multimodal framework demonstrates the value of deep learning-based approaches as a complement to handcrafted radiomics and provides a basis for more individualized treatment strategies.

Background: Neoadjuvant chemotherapy (NAC) is the important and effective approach to treat locally advanced breast cancer (LABC). The prediction of response to NAC prior to start is an efficient approach to obtaining perspective about the effectiveness of treatment. The objective of this study is to design a machine learning pipeline to predict tumor response to NAC treatment for patients with LABC using the combination of clinical features and radiomics computed tomography (CT) features. Method: A total of 858 clinical and radiomics CT features were determined for 117 patients with LABC to predict the tumor response to NAC treatment. Since the number of features is greater than the number of samples, dimensionality reduction is an indispensable step. To this end, we proposed a novel hybrid feature selection to not only select top features but also optimize the classifier hyperparameters. This hybrid feature selection has two phases. In the first phase, we applied a filter-based strategy feature selection technique using matrix rank theorem to remove all dependent and redundant features. In the second phase, we applied a genetic algorithm which coupled with the SVM classifier. The genetic algorithm determined the optimum number of features and top features. Performance of the proposed technique was assessed by balanced accuracy, accuracy, area under curve (AUC), and F1-score. This is the binary classification task to predict response to NAC. We consider three models for this study including clinical features, radiomics CT features, and a combination of clinical and radiomics CT features. Results: A total of 117 patients with LABC with a mean age of 52 ± 11 were studied in this study. Of these, 82 patients with LABC were the responder group (response to NAC) and 35 were the non-response group to chemotherapy. The best performance was obtained by the combination of clinical and CT radiomics features with Accuracy = 0.88. Conclusion: The results indicate that the combination of clinical features and CT radiomic features is an effective approach to predict response to NAC treatment for patients with LABC.

(1) Background: Response to neoadjuvant chemotherapy (NAC) is a key prognostic indicator in breast cancer. However, current response evaluation methods rely on histopathological assessment after surgery, delaying opportunities for early treatment adaptation. This study aimed to develop a machine learning model by integrating radiomic features extracted from pre-treatment, contrast-enhanced computed tomography (CT) images with baseline clinical variables to predict NAC response before therapy initiation. (2) Methods: The study investigated two categories of response: (i) pathologic complete response (pCR) versus non-pCR, and (ii) clinical response versus non-response, where clinical response was defined as a reduction in tumor size of at least 30%, encompassing both complete and partial responses. Radiomic features (n = 214) were extracted from intratumoral and peritumoral regions of pre-treatment CT images. Clinical variables (n = 7) were also incorporated to enhance predictive capability. A predictive model was developed using XGBoost algorithm, and performance was evaluated across ten independent data partitions using metrics including accuracy, precision, sensitivity, specificity, F1-score, and AUC. (3) Results: A total of 177 patients were enrolled in the study. The combined clinical-radiomic model set exhibited superior predictive performance compared to models based solely on either radiomic or clinical features. For pCR classification, integrating clinical and radiomic features produced the strongest model, achieving 82.8% accuracy with an AUC of 0.846. The clinical model alone reached 71.4% accuracy and an AUC of 0.797, while the radiomic model achieved 67.5% accuracy and an AUC of 0.615. For clinical response classification, the combined model again outperformed the individual models, achieving 71.7% accuracy with an AUC of 0.725, compared with 65.0% accuracy and an AUC of 0.666 for the clinical model, and 65.6% accuracy with an AUC of 0.615 for the radiomic model. (4) Conclusions: These results demonstrate that integrating CT radiomic features with clinical information enhances the prediction of NAC response, supporting the potential for earlier and more personalized therapeutic decision-making in breast cancer management.

The standard treatment for patients with locally advanced non-small cell lung cancer (NSCLC) is concurrent chemoradiation. However, clinical responses are heterogeneous and generally not known until after the completion of therapy. Multiple studies have investigated imaging predictors (radiomics) for different cancer histologies, but little exists for NSCLC. The objective of this study was to develop a multivariate CT-based radiomics model to a priori predict responses to definitive chemoradiation in patients with lung adenocarcinoma. Methods: Patients diagnosed with locally advanced unresectable lung adenocarcinoma who had undergone chemoradiotherapy followed by at least one dose of maintenance durvalumab were included. The PyRadiomics Python library was used to determine statistical, morphological, and textural features from normalized patient pre-treatment CT images and their wavelet-filtered versions. A nested leave-one-out cross-validation was used for model building and evaluation. Results: Fifty-seven patients formed the study cohort. The clinical stage was IIIA-C in 98% of patients. All but one received 6000–6600 cGy of radiation in 30–33 fractions. All received concurrent platinum-based chemotherapy. Based on RECIST 1.1, 20 (35%) patients were classified as responders (R) to chemoradiation and 37 (65%) patients as non-responders (NR). A three-feature model based on a KNN k = 1 machine learning classifier was found to have the best performance, achieving a recall, specificity, accuracy, balanced accuracy, precision, negative predictive value, F1-score, and area under the curve of 84%, 70%, 80%, 77%, 84%, 70%, 84%, and 0.77, respectively. Conclusions: Our results suggest that a CT-based radiomics model may be able to predict chemoradiation response for lung adenocarcinoma patients with estimated accuracies of 77–84%.

Previous work has demonstrated quantitative ultrasound (QUS) analysis techniques for extracting features and texture features from ultrasound radiofrequency data which can be used to distinguish between benign and malignant breast masses. It is desirable that there be good agreement between estimates of such features acquired using different ultrasound devices. Handheld ultrasound imaging systems are of particular interest as they are compact, relatively inexpensive, and highly portable. This study investigated the agreement between QUS parameters and texture features estimated from clinical ultrasound images of breast tumors acquired using two different ultrasound scanners: a traditional cart-based system and a wireless handheld ultrasound system. The 28 patients who participated were divided into two groups (benign and malignant). The reference phantom technique was used to produce functional estimates of the normalized power spectra and backscatter coefficient for each image. Root mean square differences of feature estimates were calculated for each cohort to quantify the level of feature variation attributable to tissue heterogeneity and differences in system imaging parameters. Cross-system statistical testing using the Mann–Whitney U test was performed on benign and malignant patient cohorts to assess the level of feature estimate agreement between systems, and the Bland–Altman method was employed to assess feature sets for systematic bias introduced by differences in imaging method. The range of p-values was 1.03 × 10⁻⁴ to 0.827 for the benign cohort and 3.03 × 10⁻¹⁰ to 0.958 for the malignant cohort. For both cohorts, all five of the primary QUS features (MBF, SS, SI, ASD, AAC) were found to be in agreement at the 5% confidence level. A total of 13 of the 20 QUS texture features (65%) were determined to exhibit statistically significant differences in the sample medians of estimates between systems at the 5% confidence level, with the remaining 7 texture features being in agreement. The results showed a comparable magnitude of feature variation between tissue heterogeneity and system effects, as well as a moderate level of statistical agreement between feature sets.