Machine Learning and Deep Learning in Oncologic Imaging: Potential Hurdles, Opportunities for Improvement, and Solutions-Abdominal Imagers' Perspective.

ABSTRACT: The applications of machine learning in clinical radiology practice and in particular oncologic imaging practice are steadily evolving. However, there are several potential hurdles for widespread implementation of machine learning in oncologic imaging, including the lack of availability of a large number of annotated data sets and lack of use of consistent methodology and terminology for reporting the findings observed on the staging and follow-up imaging studies that apply to a wide spectrum of solid tumors. This short review discusses some potential hurdles to the implementation of machine learning in oncologic imaging, opportunities for improvement, and potential solutions that can facilitate robust machine learning from the vast number of radiology reports and annotations generated by the dictating radiologists.

Temporal evolution of metastatic disease: part II-a novel proposal for subcategorization of metastatic disease from non-neural solid tumors with diverse histologies and locations.

Tumor spread is a continuous process and metastases can further disseminate. Currently, metastatic disease from most primary tumors is subcategorized as M0 if absent and M1 if present. However, metastatic disease in different locations may have different prognostic implications, even if it is from the same primary tumor. The current staging systems for metastatic disease have not evolved to match our understanding of the disease's evolution or the evolving treatment paradigms. Primary tumor-specific subcategorization of metastatic disease is currently available for a few tumors, but not all of them imply further remote spread of tumor, similar to tumor (T) and N (node) subcategorizations of the TNM staging, nor are they applicable to wide spectrum of other tumors. In this era of precision medicine, tumor-type agnostic therapies based on common biomarkers rather than primary tumor sites are emerging, but a subcategorization system applicable to metastatic disease from diverse primary tumor locations and with diverse histologies is not available. In this article, we discuss the need to further classify the metastatic disease and present a subcategorization applicable to metastatic disease from non-neural solid tumors from different primary tumor sites and with different histologies, which is based on the temporal spread of metastatic disease. Our proposed subcategorization scheme for metastatic disease into M0, M1, M2 and M3, is universally applicable to a diverse spectrum of non-neural solid tumors, and increasing M subcategorization represents further remote spread of tumor.

Temporal evolution of metastatic disease: part I-an in-depth review of the evolution of metastatic disease across diverse spectrum of non-neural solid tumors on serial oncologic imaging studies and relevant practical applications.

With improved survival rates of patients with metastatic disease due to continuously evolving multimodality treatment options, radiologists are increasingly interpreting imaging studies from patients with protracted metastatic disease. It is thus crucial for radiologists to have an in-depth understanding of the temporal evolution of metastatic spread and the accompanying findings on imaging studies, to provide accurate interpretation that supports optimal management. A general overview of the evolution of cancer spread on serial imaging studies and common pathways of tumor spread across multiple tumor types and tumor locations is not readily available in radiology literature. The key common pathways of tumor spread across diverse spectrum of tumors relevant to radiologists are summarized in a logical schematic approach which focusses on aiding radiologists to understand the pathways of spread resulting in current sites of metastatic disease involvement and then to potentially predict future sites of metastatic involvement. This article also summarizes the practical applications of this knowledge to the routine oncologic imaging interpretation.