Deep Learning for Bone Metastasis Localisation in Nuclear Imaging data of Breast Cancer Patients

Abstract

Bone scintigraphy is a popular method for the diagnosis of bone metastasis that typically occurs when cancer cells from the primary tumor relocate to the bone. In bone scintigraphy, the whole patient's body is scanned and the generated bone scan visualization provides a valuable source of information for the evaluation of various bone-related pathologies, including bone inflammation and fractures, nonmalignant bone lesions, bone infections, or even the spread of cancer to the bone. ?n particular, bone cancer is among the most frequently appeared diseases to patients suffering from metastatic cancer such as breast cancer patients. However, hot spots in bone scans indicating inflammations or cancer metastasis can be misleading. Accurate detection of pathological hot spots can be a very challenging procedure, with the experience of clinicians playing a critical role in the interpretation of the images. Artificial intelligence has emerged as a key enabler in the interpretation of medical imaging being able to model the aforementioned uncertainties and providing a reliable automated solution. So far, a number of convolutional neural networks (CNN)-based techniques have been proposed in the recent literature coping with the problem of bone metastasis classification. To the best of our knowledge, localization of pathological and degenerative hot spots in scintigraphy images is a scientific area that has not been explored. This paper contributes to the first ever deployment of advanced deep learning networks for bone metastasis localization in nuclear imaging data of breast cancer patients. The methodology relies on the latest advances of object detection via the use of two powerful and recent models (scaled YOLO v4 and Detectron2). The efficacy of the proposed methodology was demonstrated utilizing an extensive experimentation setup. The proposed methodology demonstrates unique potential in bone metastasis localization therefore facilitating the clinical interpretation of bone scintigraphy scans. © 2021 IEEE.