Training Deep CNN's to Detect Prostate Cancer Lesion with Small Training Data

Prostate cancer is one of the most common cancers among men. Fatality due to prostate cancer can be reduced if the cancerous lesions are spotted in the early stage. Although deep neural networks showed superb performances in many healthcare tasks, in some tasks they were still under-explored. This work focused on detecting prostate cancer lesions in MRI images where the size of the dataset is comparatively small. We studied three prominent Deep Convolutional Neural Networks for Object Detection on how they perform on the task and we compared them with the baseline algorithm which is an off-the-shelf Computer Vision in the Cloud service. Furthermore, we explored impact of incorporating additional data from publicly available datasets into the training data. Our experiment showed that the EfficientDet model outperformed the baseline model, YOLOv4 model, and YOLOv5 model by achieving 52.63% accuracy when training with a small training data. In addition, we found that only the YOLOv4 model had significantly higher precision when incorporating publicly available data into the training data.