“Small-data” Patch-wise Multi-dimensional Output Deep-learning for Rare Cancer Diagnosis in MRI under Limited Sample-size Situation

Soft-tissue sarcomas are a group of heterogeneous rare cancers that are challenging in diagnosis by non-expert doctors. Deep learning is a promising approach in artificial intelligence (AI)-aided diagnosis when "big data" (e.g., from 10,000 to 100,000 cases) is available. As such, building a deep-learning-based AI system for rare cancers is very difficult due to the limited sample-size nature of rare cancer. In this study, we proposed patch-wise compact neural networks to build a rare cancer diagnosis model with a small sample-size dataset. Our model predicts a probability map of malignancy instead of a single possibility value of malignancy or a category of a malignant tumor. Our model was trained and tested with a database containing 96 benign tumors and 50 malignant rare tumors in magnetic resonance imaging (MRI). Our proposed model achieved an area-under-the-curve (AUC) performance of 0.78 in the classification of malignant and benign tumors, statistically significantly higher than those (0.64-0.73) by conventional radiomics and convolutional neural network (CNN)-based models. Furthermore, we provided the visualization of the probability maps from the proposed model, which directly explained how the model made predictions from images.