Software Solution for Clinical Management of Musculoskeletal Tumors - PROJECT SUMMARY/ABSTRACT
Bone lesions are frequently encountered during every day clinical practice. Benign bone lesions, such as a
small enchondroma, can be left alone and are unlikely to impact the patient during their lifetime. However, a
malignant bone lesion, such as an osteosarcoma, will require a biopsy and surgical resection. Determining
which lesions require treatment and which can be left alone can be a daunting process. Whether a bone lesion
requires a biopsy depends on both clinical- and imaging-based factors(1–3). Advanced patient age, presence
of pain, and history of prior malignancy can influence the need for biopsy. For imaging, various lesion-based
parameters such as location, matrix, tumor margin, presence of soft tissue component, and periosteal reaction
can help determine whether the lesion is aggressive or non-aggressive, with aggressive lesions needing a
biopsy. Aggressive lesions are more likely to represent a malignancy, althought there are some benign
processes that can have an aggressive imaging appearance (i.e osteomyelitis, fractures).
Misdiagnoses of malignant tumors as benign prevents needed treatment from occurring; however, benign
lesions should not be unnecessarily biopsied, as this can lead to unneeded tests, biopsy complications,
increased health care costs, and patient anxiety. Currently, the decision to biopsy or not is made by the clinician,
considering the clinical- and imaging-based factors, which can be very subjective. Studies have shown that
misdiagnosis is higher if these cases are not discussed under multidisciplinary review with input from an
orthopedic oncologist, radiologist, and pathologist. Also, if the imaging studies are not interpreted by
subspecialty trained musculoskeletal (MSK) radiologists, reading discrepancy of up to 28% can occur. Moreover,
a recent study by Zamora et al. showed that there is poor inter-observer agreement amongst experienced
orthopedic oncologists for distinguishing enchondromas and chondrosarcomas, a common clinical dilemma.
Therefore, we propose to develop a method to analyze the radiologic studies directly, extract important
lesion-based features of the bone tumors and auto-classify the lesions as non-aggressive or aggressive. By
using CT scans from 200 biopsied bone lesions, utilizing a deep learning approach to extract image features and
access to patients' clinical-based factors, we will develop a machine learning tool to differentiate between non-
aggressive and aggressive tumors and compare the results to definitive histologic confirmation of disease. We
hypothesize that the proposed machine learning based software will classify aggressive vs. non-
aggressive lesions as accurately as definitive histologic confirmation of disease state. The aims of the
study are to: 1) Develope bone and lesion segmentation and bone-lesion feature extraction software tools for
physician classification; and 2) Develope a software tool to auto-classify femur lesions as either aggressive or
non-aggressive.
