# Research on the application of a multi-model cascaded deep learning framework in the pathological diagnosis of osteosarcoma

## TL;DR

This paper introduces a deep learning framework to improve the accuracy of diagnosing and evaluating osteosarcoma, a bone cancer in adolescents.

## Contribution

A novel multi-model cascaded deep learning framework using a Vision Mamba model for precise osteosarcoma diagnosis and evaluation.

## Key findings

- The model achieved Dice coefficients of 0.83 or higher in tumor and matrix segmentation tasks.
- Performance metrics like AUC, sensitivity, and specificity exceeded 90% in necrosis classification and subtype detection.
- The framework shows high potential for clinical application in improving osteosarcoma diagnosis precision.

## Abstract

Osteosarcoma is the most common malignant tumor of bone tissue in adolescents, and precise pathological diagnosis is the primary foundation for establishing the most effective treatment plan. The pathological evaluation of tumor necrosis after chemotherapy is crucial for assessing therapeutic efficacy in osteosarcoma patients. However, pathologists often face several challenges during the diagnosis and evaluation process.

To address these needs, we designed and developed a multi-model cascaded deep learning framework utilizing an advanced Vision Mamba (ViM) model as the core network architecture. The study employed one of the most comprehensive osteosarcoma datasets, sourced from: (1) real-world data from 68 osteosarcoma patients collected at Chongqing General Hospital, and (2) publicly available osteosarcoma assessment data from the University of Texas Southwestern/UT Dallas. Pathological images were annotated using the Palgo pathology image artificial intelligence self-training platform according to algorithm requirements. A triple verification mechanism of annotation, review, and archiving was implemented, and Palgo’s integrated interactive algorithm correction mechanism was used to continuously refine the data annotation process.

The model demonstrated Dice coefficient values of 0.83 or higher in tumor segmentation, osteosarcoma osteoid matrix segmentation, necrotic area segmentation, lung metastatic tumor segmentation, and lung metastatic osteoid matrix segmentation. For necrosis classification, overall osteosarcoma subtypes, and localized osteosarcoma subtypes, the area under the receiver operating characteristics curve (AUC), sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) all exceeded 90%. The proposed model exhibited excellent performance, indicating high potential for future clinical application in osteosarcoma patients. This framework shows promise for enhancing the precision and efficiency of pathological diagnosis and evaluation in osteosarcoma management.