Diagnosing low-grade central osteosarcoma using a neural network mathematical model: a case report and review

BACKGROUND: Diagnosing low-grade central osteosarcoma is associated with a significant challenge because, according to radiological and histological findings, the condition closely resembles various benign lesions, most commonly being misdiagnosed as fibrous dysplasia. Convolutional neural network-based mathematical models have been successfully applied for the automated analysis of digital histopathological images, including tumor classification, regions of interest segmentation, and identification of morphological features of malignancy.

CASE DESCRIPTION: This paper presents a clinical case of a 33-year-old female patient in whom, following a pathologic fracture of the femoral diaphysis, the lesion was long misinterpreted as fibrous dysplasia. Upon re-evaluation of histological slides and repeat biopsy at the N.N. Priorov National Medical Research Center of Traumatology and Orthopedics, the diagnosis of low-grade central osteosarcoma with areas of dedifferentiation and formation of high-grade osteosarcoma foci was established. For additional diagnostic confirmation, a convolutional neural network (ResNet-101)-based mathematical model previously developed by the authors for automated detection of pathologic mitoses on digital histopathological images was applied. The model analyzed scanned slides (Leica Aperio CS2, ×400), identifying several structures with a high probability of pathologic mitoses (maximum confidence scores: 99% and 92%), consistent with the conclusions of two experienced pathologists, thereby confirming the malignant nature of the lesion.

CONCLUSION: This paper presents a clinicopathological and radiologic description of the condition, discusses diagnostic challenges and similarities with fibrous dysplasia and other benign lesions, and evaluates the potential and limitations of artificial intelligence techniques in pathology for rare low-mitotic tumors. Emphasis is placed on the role of neural network analysis as an auxiliary tool for improving reproducibility and sensitivity of mitosis detection, the need for multicenter model validation, and the implementation of stain normalization and interpretability of results for clinical application.