A Self-Selective Spatial–Semantic Deep Prediction Network for Automated Detection and Quantitative Severity Assessment of Bone Sarcoma

Background

Bone Sarcoma is a most common primary bone tumor with complex spatial morphology, heterogeneous tissue

constitution, which complicates the process of localizing the tumor very well and objectively determining their

severity in histopathological images. Existing systems of deep learning are predominantly performing single tasks

such as classification or segmentation without a cohesive system capable of simultaneously measuring multi-scale

spatial structure and high-level semantic malignancy features to inform severity-sensitive diagnosis.

Objective

To deal with this shortcoming, this research will propose a Self-Selective Spatial-Semantic Deep Prediction

Network (S-DPN) that can be used to automatically detect bone sarcoma and the severity thereof in a quantitative

manner.

Materials and Methods

The architecture has parallel spatial and semantic encoders to achieve structural features, which are related to tumor

shape, irregularity of the periphery and the pattern around them, which are linked with malignancy of the tissues.

Self-selective attention process is an adaptive process whereby there is stressing of clinically relevant areas and

inhibition of background noises followed by weighted spatial and semantic fusion. The combined features not only

aid in the appropriate localization of the tumor, but also in the determination of the severity of the tumor in

percentage of the affected bone tissue that can be interpreted as a clinical value. S-DPN adequately addresses the

limitations in the prediction of tumor presence and a quantitative measure of the severity of the tumor by jointly

optimizing both detection and severity estimation in an end-to-end learning system.

Results

The experimental findings on publicly accessible bone sarcoma histopathology dataset indicate that the proposed

model has accuracy of 98.7% and it is superior to the conventional deep architectures.

Conclusion

The S-DPN provides clinically informative severity-conscious, high-quality bone sarcoma determination to support

correct diagnosis and treatment planning.