Diagnosis of low-grade central osteosarcoma using a neural network model. Case report and literature review

BACKGROUND: The diagnosis of low‑grade central osteosarcoma poses a significant diagnostic challenge because its radiologic and histologic features substantially overlap with various benign processes, most commonly leading to misdiagnosis as fibrous dysplasia. Convolutional neural network–based mathematical models are successfully applied for automated analysis of digitized histological images, including tumor classification, segmentation of regions of interest, and identification of morphological features of malignancy. One of the most promising directions is automated detection of mitoses — a key indicator of tumor proliferative activity with important diagnostic and prognostic value.

CLINICAL CASES DESCRIPTION: We report a clinical case of a 33-year-old female whose lesion was long and erroneously interpreted as fibrous dysplasia after a pathological fracture of the femoral diaphysis. Review of histological specimens and repeat biopsy at the N.N. Priorov National Medical Research Center for Traumatology and Orthopedics established a diagnosis of low‑grade central osteosarcoma with areas of dedifferentiation and foci of high‑grade osteosarcoma. For diagnostic support, a mathematical model based on a convolutional neural network (ResNet-101), previously developed by the authors for automated detection of pathological mitoses in digitized histological images, was used. The model analyzed scanned slides (Leica Aperio CS2, ×400) and identified several objects with high probabilities of being pathological mitoses (maximum probability scores 99% and 92%), findings that corresponded to the assessments of two experienced pathologists and corroborated the malignant nature of the lesion.

CONCLUSION: A clinicopathological and radiological description of the disease is presented. Diagnostic difficulties and similarities with fibrous dysplasia and other benign processes are discussed, as well as the potential and limitations of artificial intelligence methods in pathology for rare tumors with low mitotic activity. Emphasis is placed on the role of neural‑network analysis as an adjunct tool to improve reproducibility and sensitivity of mitosis detection, and on the need for multicenter model validation, implementation of color‑normalization methods, and result interpretability prior to clinical deployment.