Mask R-CNN Machine Learning Based Pancreatic Lesion Detection and Categorization: With An Open-source Tool For Real World Use

Abstract

Interest in applying machine learning (ML) in medical research has exploded over the past decade, especially in oncology and medical imaging. The objective of ML, particularly machine vision in medical imaging, is to solve lesion identification, segmentation, and classification problems. Although numerous studies have developed ML models that can be beneficial in various clinical settings, more often than not, there is no easy way to apply the models in practice without technical knowledge. To address this issue, in the process of developing an ML model to detect and categorize pancreatic lesions as either cystic or solid, we developed the PCP2 Toolkit. PCP2 is an open-source tool for training and using Mask R-CNN-based ML vision models. PCP2’s core ML functionality is implemented in Python and the user interface in JavaScript. Here, we discuss the architecture, workflow, and application of PCP2 and demonstrate its use in training a model to classify pancreatic lesions on Computed Tomography (CT) images. The model achieved a mean average precision (mAP) of 61.2% and a mean average recall (mAR) of 65.1%. The mAP at Intersection over Union (IoU) thresholds of 50% and 75% were 80.6% and 65.1%, respectively. The source code, documentation, and trained model are available at nbcatalyst.com. Our objective is to provide an easy-to-use, customizable, and extendable tool for training machine vision models and using the trained model in the real world.