2026-10-13T08:50:56Z https://api.figshare.com/v2/oai

oai:figshare.com:article/32640213 2026-10-01T16:33:13Z portal_1172 item_type_8 month_year_10_2026

Deep learning breast cancer radiology anomaly detection explainability using statistical fault localization Jonathan Armstrong (13148871) PUREID: 613870496 medical imaging mammography machine learning explainability fuzzing DeepCover The World Health Organisation (WHO) states that cancer is the leading cause of death in the in the world accounting for an estimated 9.6 million deaths in 2018 [1] Breast and Colorectal being the second and third most common respectively [1] it also estimates that 30-50% of cancer deaths can be avoided by following current prevention strategies. Early detection is also key to avoiding the more serious cases [1].<br><br>The aim of the project is the design, implementation, and evaluation of deep learning neural networks for the detection of anomalies in breast mammograms in line with current state-of-the-art. The best algorithm will then be analysed using a fuzzing technique developed by Sun et. al. [2] which uses a Statistical Fault Localization approach to define the minimal input features capable of influencing the final algorithm decision, this approach has not been used as yet in greyscale or medical images such as mammograms and should offer unique insights into how the algorithm detects these anomalies.<br><br>This research will extensively test a number of algorithms and hyperparameters for mammogram detection, once the most appropriate algorithm/parameter set is defined, try to understand, and explain the outputs, through an extensive study of the misclassified patches, and a sliding window classifier approach on whole images. Once this is completed and an understanding is gained of the main influences on the algorithms, then we will as already mentioned use a novel fuzzing technique developed by Sun et. al. [2] to progress past patch-wise classification and define the minimal input features capable of influencing the final algorithm decision. For the mammograms initially the aim was to create our own dataset but since the project is alongside a company the focus changed and we used the open dataset MIAS [3] containing 322 annotated images.<br><br> 2026-10-01T16:33:13Z Text Thesis 10.17034/32640213.v1 https://figshare.com/articles/thesis/Deep_learning_breast_cancer_radiology_anomaly_detection_explainability_using_statistical_fault_localization/32640213 All Rights Reserved