A Robust Lightweight Neural Network Architecture for Real-Time Point-of-Care (POC) Ultrasound Interpretation using Medical Image Evaluation

The proposed paper presents a strong and slim neural network model specially designed to operate in point-of-care (POC) ultrasound interpretation in real-time. The proposed model is meant to be deployed in the low-resource clinical environment, as it combines a MobileNetV3-based backbone with self-supervised pretraining and multi-task learning to classify diagnostic categories and label anatomical regions at the same time. Channel pruning and quantization are used to optimize the system, and it can be converted into TensorRT to execute efficient edge inference on such devices as Jetson Nano and Raspberry Pi. Vast experimentation on various ultrasound datasets, such as abdominal, cardiac, obstetric, and musculoskeletal scans, shows better results on various important metrics. On edge devices, the proposed model had a classification accuracy of 95.4 along with a Dice segmentation score of 93.4 and inference latency of less than 30 milliseconds; which was better than traditional CNN baselines. Its findings confirm its real-time nature, diagnostic soundness and its portability as POC ultrasound. The clinical validation of this model highlights the model as useful in frontline workers in health care as well as in mobile diagnostic teams especially in remote and underserved areas. The future research will be on the addition of attention-based visual explainability, cross-device generalization, and federated learning to achieve privacy-preserving adaptation. In general, the framework offers a realistic and scalable approach to democratizing access to real-time and resource-constrained medical image interpretation that is driven by AI.