The long-term goal of this project is to develop a unified software package for sharing image analysis and machine learning tools to improve the accuracy and efficiency of cancer diagnosis, thus aiding in improving the quality of both cancer research and clinical practice. Since diagnostic errors can cause altered treatment recommendations and significant patient harm, tools that semi-automate the diagnostic process to improve efficiency, reliability and accuracy are thus needed by both cancer researchers and clinical pathologists. Our work will produce a set of machine-learning tools to identify regions of interest (ROI), classify these regions into the full spectrum of diagnostic categories (ranging from benign, to pre-invasive/risk lesions, to invasive cancer), and culminate in the development of a unified software package for cancer diagnosis that can be shared among cancer researchers and clinicians. Our aims are: Aim 1: Regions of Interest. Produce 1a) a ROI finder classifier and associated tools for use by researchers or pathologists for automatic identification of potential ROIs on whole slide images of breast biopsy slides and 1b) a ROI analysis classifier and associated tools that can point out image regions that tend to cause misdiagnosis and produce suitable warnings as to why such regions may either be distractors or indicate cancer. Aim 2: Diagnosis. Produce a diagnostic classifier and associated tools that can not only suggest the potential diagnosis of a whole slide image, but can also produce the reasons for the diagnosis in terms of regions on the image, their color, their texture, and their structure. Aim 3: Dissemination. Develop a unified software package containing this suite of tools, so they can be easily shared and provided (standalone and through the existing PIIP platform) to both cancer researchers and clinical pathologists. In addition to specific classifiers for breast cancer research, we will provide the methodology to train related classifiers for other biopsy- diagnosed cancers, such as melanoma, prostate, lung, and colon cancer. Our highly innovative diagnostic tools will include our state-of-the-art (2017) feature extraction and machine learning methods as opposed to the generic image management tools of the current ITCR projects. Our classifiers were trained on a unique data set acquired in a carefully designed breast cancer research study, so they are immediately useful in the breast cancer domain, but designed to be easily retrained for other data sets and cancers. The addition of our unique multi-disciplinary team and these tools to the ITCR program will be an important aid to cancer research and will improve the process of diagnosis for clinicians.

Quality medical care requires an accurate diagnosis. Diagnostic errors lead to lower quality research and can lead to reduced quality of care and inappropriate treatment decisions that result in harm to patients. Our team will develop a unified software package for sharing image analysis and machine learning tools to improve the accuracy and efficiency of cancer diagnosis and will share these tools with cancer researchers and pathologists on a publicly available platform.