Mohamed Harmanani

Many medical AI models perform unevenly across patient groups because they learn shortcuts from biased data. These hidden biases make models less reliable and less fair in real-world use. We introduce DPE-Former, a model that combines prototype-based learning with transformer attention. The system trains several complementary classifiers on balanced subsets of data, each capturing different aspects of the population. A transformer module then learns how to combine its outputs in an adaptive way, helping the model make more balanced decisions across unseen or minority groups. DPE-Former achieved higher accuracy on underrepresented groups and more consistent performance overall compared to standard training methods.

Medical foundation models (FMs) offer a path to build high-performance diagnostic systems. However, their application to prostate cancer detection from micro-ultrasound (μUS) remains untested in clinical settings. We present ProstNFound+, an adaptation of FMs for μUS-based PCa detection, and report the first prospective validation of such a system. The model demonstrates strong agreement with biopsy-confirmed findings and surpasses conventional scoring protocols.

Deep learning holds significant promise for enhancing real-time ultrasound-based prostate biopsy guidance through precise and effective tissue characterization. Despite recent advancements, prostate cancer (PCa) detection using ultrasound imaging still faces two critical challenges: (i) limited sensitivity to subtle tissue variations essential for detecting clinically significant disease, and (ii) weak and noisy labeling resulting from reliance on coarse annotations in histopathological reports. To address these issues, we introduce ProTeUS, an innovative spatio-temporal framework that integrates clinical metadata with comprehensive spatial and temporal ultrasound features extracted by a foundation model. Our method includes a novel hybrid, cancer involvement-aware loss function designed to enhance resilience against label noise and effectively learn distinct PCa signatures. Furthermore, we employ a progressive training strategy that initially prioritizes high-involvement cases and gradually incorporates lower-involvement samples. These advancements significantly improve the model’s robustness to noise and mitigate the limitations posed by weak labels, achieving state-of-the-art PCa detection performance with an AUROC of 86.9%.

The subpopulation shift — a mismatch in subgroup distribution between training and deployment — can significantly degrade performance. We propose Diverse Prototypical Ensembles, which encourage classifier diversity to improve worst-group accuracy. Evaluated on nine datasets, our method outperforms prior state-of-the-art approaches.

We introduce Cinepro, a novel framework that strengthens foundation models' ability to localize PCa in ultrasound cineloops. Cinepro adapts robust training by integrating the proportion of cancer tissue reported by pathology in a biopsy core into its loss function to address label noise. It leverages temporal data across multiple frames to apply robust augmentations. Cinepro achieves an AUROC of 77.1% and balanced accuracy of 71.9%, surpassing current benchmarks.

While deep learning methods have shown great promise in improving the effectiveness of prostate cancer (PCa) diagnosis by detecting suspicious lesions from trans-rectal ultrasound (TRUS), they must overcome multiple simultaneous challenges. There is high heterogeneity in tissue appearance, significant class imbalance in favor of benign examples, and scarcity in the number and quality of ground truth annotations available to train models. Failure to address even a single one of these problems can result in unacceptable clinical outcomes. We propose TRUSWorthy, a carefully designed, tuned, and integrated system for reliable PCa detection. Our pipeline integrates self-supervised learning, multiple-instance learning aggregation using transformers, random-undersampled boosting and ensembling: these address label scarcity, weak labels, class imbalance, and overconfidence, respectively. We train and rigorously evaluate our method using a large, multi-center dataset of micro-ultrasound data. Our method outperforms previous state-of-the-art deep learning methods in terms of accuracy and uncertainty calibration, with AUROC and balanced accuracy scores of 79.9% and 71.5%, respectively. On the top 20% of predictions with the highest confidence, we can achieve a balanced accuracy of up to 91%. The success of TRUSWorthy demonstrates the potential of integrated deep learning solutions to meet clinical needs in a highly challenging deployment setting, and is a significant step towards creating a trustworthy system for computer-assisted PCa diagnosis.

This multi-center study evaluates PCa ultrasound detection under realistic data heterogeneity. It highlights the importance of uncertainty calibration for clinical decision support and establishes strong benchmarks for future research.

This study benchmarks ViT architectures for ROI and multi-scale prostate ultrasound classification, comparing them to CNNs. A multi-objective learning strategy achieves 77.9% AUROC with significant gains in sensitivity and specificity.