LiteXrayNet: Quantum-Inspired Deep Learning Framework for Scalable Pneumonia Diagnosis

TMLR Paper5796 Authors

Abstract: Pneumonia is a serious health problem affecting the world and affecting heavily low-resource areas, where timely diagnostic facilities are vital. This paper, in turn, presents liteXrayNet, an advanced convolutional neural network (CNN) that is tailored specifically to detect pneumonia on chest radiographs with high accuracy and is designed to run under conditions of limited computer resources. This network structure uses the inverted residual MBConv blocks of MobileNetV3 that can help extract features effectively, a quantum-inspired phase shift layer that can be used to enhance the detection of complex patterns, and a simplified recognizer, which will guarantee strong binary classification. With 179,646 trainable parameters, liteXrayNet achieves a test-level accuracy of 97%, has a small model size of 0.7 MB, and inference latency of 0.60 ms/sample, liteXrayNet can achieve diagnostic accuracy in real time on resource-constrained systems. The model has minimal computing requirements with little impact on diagnostic quality achieved through integrating depthwise separable convolutions, hard-swish activations and quantum-inspired feature modulation. The liteXrayNet has been demonstrated to be a efficient solution to scalable, point-of-care pneumonia diagnosis, allowing significantly more people to access and obtain healthcare and undo disparities by diagnosis in underserved populations globally, due to its lightweight construction and high diagnostic accuracy.

Submission Length: Long submission (more than 12 pages of main content)

Assigned Action Editor: ~Stanislaw_Kamil_Jastrzebski1

Submission Number: 5796