Towards Interpretable Feature Maps for Visualizing Prostate Cancer in Ultrasound Data

Abstract

Prostate cancer is the second most common cancer worldwide, but has a high long-term survival rate if detected early. The clinical standard for definitive diagnosis is through histopathological analysis of prostate tissue, obtained during trans-rectal ultrasound-guided core biopsy. Conventional ultrasound (US) has a low sensitivity, so prostate biopsies are extracted systematically from pre-defined anatomical locations. As a result, this procedure has a high false negative rate. A targeted biopsy method, where US would be used for detection in addition to navigation, would be beneficial as this modality is safe, inexpensive, and accessible to clinicians. Micro-US has recently been proposed to improve US-based detection of prostate cancer. It has a higher frequency than conventional US and can visualize histological architectures associated with prostate cancer at a much higher resolution. Deep learning methods have also been proposed for prostate cancer detection with promising results. However, the underlying decisions these models make to arrive at a prediction is unclear and the interpretability of these models is required to foster clinical trust. To contribute to the overarching goal of creating an accurate US-based prostate cancer detection method, we propose using deep learning to create interpretable feature maps for visualizing prostate cancer in micro-US data. To do this, we extracted high-dimensional features from a deep learning model trained on multi-center data to classify cancer in micro-US images. We then conducted experiments using different dimensionality reduction techniques to visualize the classifier latent space in a way that is interpretable to humans. The resulting feature maps were compared and observed to analyze patterns occurring within the latent space. After comparing the feature maps resulting from the different dimensionality reduction techniques, it was determined that the low-dimensional embeddings generated by UMAP could successfully distinguish between benign and cancer micro-US features. Most significantly, the UMAP embedding identified distinct clusters associated with Gleason grades, a measure of prostate cancer severity, despite the original classifier being trained binary labels. These maps suggest that the model may be learning histopathologically informative representations in the latent space, similar to those observed by pathologists when diagnosing prostate cancer.