Machine learning models have been successfully applied for analysis of skin images. However, due to the black box nature of such deep learning models, it is difficult to understand their underlying reasoning. This prevents a human from validating whether the model is right for the right reasons. Spurious correlations and other biases in data can cause a model to base its predictions on such artefacts rather than on the true relevant information. These learned shortcuts can in turn cause incorrect performance estimates and can result in unexpected outcomes when the model is applied in clinical practice. This study presents a method to detect and quantify this shortcut learning in trained classifiers for skin cancer diagnosis, since it is known that dermoscopy images can contain artefacts. Specifically, we train a standard VGG16-based skin cancer classifier on the public ISIC dataset, for which colour calibration charts (elliptical, coloured patches) occur only in benign images and not in malignant ones. Our methodology artificially inserts those patches and uses inpainting to automatically remove patches from images to assess the changes in predictions. We find that our standard classifier partly bases its predictions of benign images on the presence of such a coloured patch. More importantly, by artificially inserting coloured patches into malignant images, we show that shortcut learning results in a significant increase in misdiagnoses, making the classifier unreliable when used in clinical practice. With our results, we, therefore, want to increase awareness of the risks of using black box machine learning models trained on potentially biased datasets. Finally, we present a model-agnostic method to neutralise shortcut learning by removing the bias in the training dataset by exchanging coloured patches with benign skin tissue using image inpainting and re-training the classifier on this de-biased dataset.

Tools based on deep learning models have been created in recent years to aid radiologists in the diagnosis of breast cancer from mammograms. However, the datasets used to train these models may suffer from class imbalance, i.e., there are often fewer malignant samples than benign or healthy cases, which can bias the model towards the healthy class. In this study, we systematically evaluate several popular techniques to deal with this class imbalance, namely, class weighting, over-sampling, and under-sampling, as well as a synthetic lesion generation approach to increase the number of malignant samples. These techniques are applied when training on three diverse Full-Field Digital Mammography datasets, and tested on in-distribution and out-of-distribution samples. The experiments show that a greater imbalance is associated with a greater bias towards the majority class, which can be counteracted by any of the standard class imbalance techniques. On the other hand, these methods provide no benefit to model performance with respect to Area Under the Curve of the Recall Operating Characteristic (AUC-ROC), and indeed under-sampling leads to a reduction of 0.066 in AUC in the case of a 19:1 benign to malignant imbalance. Our synthetic lesion methodology leads to better performance in most cases, with increases of up to 0.07 in AUC on out-of-distribution test sets over the next best experiment.

Tools based on deep learning models have been created in recent years to aid radiologists in the diagnosis of breast cancer from mammograms. However, the datasets used to train these models may suffer from class imbalance, i.e., there are often fewer malignant samples than benign or healthy, which can bias the model towards the healthy class. In this study, we systematically evaluate several popular techniques to deal with this class imbalance, namely, class weighting, over-sampling, and under-sampling, as well as a synthetic lesion generation approach to increase the number of malignant samples. These techniques are applied when training on three diverse Full-Field Digital Mammography datasets, and tested on in-distribution and out-of-distribution samples. The experiments show that a greater imbalance is associated with a greater bias towards the majority class, which can be counteracted by any of the standard class imbalance techniques. On the other hand, these methods provide no benefit to model performance with respect to Area Under the Curve of the Recall Operating Characteristic (AUC-ROC), and indeed under-sampling leads to a reduction of 0.066 in AUC in the case of a 19:1 benign to malignant imbalance. Our synthetic lesion methodology leads to better performance in most cases, with increases of up to 0.07 in AUC on out-of-distribution test sets over the next best experiment.

Detecting multiple sclerosis lesions in spinal cord MRI is a critical but complex task for radiologists and neurologists. While deep learning models have shown promise for this task, training these models requires large manually annotated datasets, which are time consuming and expensive to create. To reduce the annotation burden, this study proposes LesionSCynth, a parametric framework for synthesising hyperintense lesions in spinal cord MRI. Based on analysis of the intensity distribution of real lesions in sagittal T2-weighted acquisitions, LesionSCynth generates realistic synthetic lesions that can augment small annotated datasets. Segmentation models trained on a combination of these synthetic lesions and 17 real acquisitions achieved notably better performance than those trained on the real lesions alone (0.52 vs. 0.46 FROC, respectively). Notably, models trained with LesionSCynth lesions achieved similar detection performance to a model trained on eight times more real lesions (0.55 FROC for both). Moreover, LesionSCynth outperformed two leading lesion synthesis methods, LesionMix (0.43) and CarveMix (0.41), in low-data regimes. These findings position LesionSCynth as a practical and effective solution for reducing the annotation burden while improving multiple sclerosis (MS) lesion detection and segmentation in spinal cord MRI.

Détection et segmentation automatiques, précises et robustes des lésions de sclérose en plaques dans les IRM de la moelle épinière: Cette thèse porte sur l'amélioration de la segmentation et de la détection des lésions de sclérose en plaques (SEP) dans l'IRM de la moelle épinière, une tâche important et complexe en pratique clinique. Une étude de la segmentation manuelle des lésions a mis en évidence une variabilité importante entre les experts, soulignant la nécessité de méthodes automatisées robustes et sensibles. Pour y répondre, nous avons étudié des stratégies de fusion pour combiner l'information de plusieurs images acquises avec des séquences IRM différentes. La fusion tardive s’est révélée particulièrement efficace, surpassant nettement les modèles basés sur une seule séquence. Enfin, la thèse introduit LesionSCynth, un cadre simple et efficace pour la synthèse de lésions réalistes de la moelle épinière, ce qui permet d'améliorer les performances dans des contextes de données limitées et de réduire la dépendance aux annotations manuelles. Globalement, les contributions de cette thèse font progresser le développement d’outils de segmentation fiables, généralisables et économes en données pour l’analyse des lésions de la moelle épinière dans la SEP.