Abstract
Liver diseases are a leading cause of death worldwide, particularly in sub-Saharan Africa where infectious diseases like hepatitis B and schistosomiasis are prevalent. Point-of-care ultrasound (POCUS) is a vital diagnostic tool in low-resource settings due to its portability and low cost. However, automated classification models require accurate identification of liver tissue, which is often difficult due to noisy images and ill-defined boundaries. We present a fully automated pipeline that identifies liver tissue by using the diaphragm as an anatomical anchor. By first detecting and segmenting the diaphragm, the system infers the liver's position. This approach achieves an Intersection over Union (IoU) of 0.83, significantly outperforming direct liver segmentation (0.62 IoU) and demonstrating improved clinical utility.
Introduction
In regions like rural Uganda, schistosomal periportal fibrosis (PPF) is a major health concern. While POCUS is the gold standard for diagnosis, the resulting scans are often lower in quality than high-end clinical machines. For AI models to stage liver disease, they must first isolate the liver tissue (the Region of Interest). Direct segmentation often fails because liver boundaries are "fuzzy" in POCUS scans, and models frequently mistake internal vessel walls for the liver's outer boundary. This is particularly problematic for PPF, which develops specifically along those vessel walls.
Methods
The proposed pipeline avoids the "fuzzy boundary" problem by focusing on the diaphragm—a highly echogenic structure that appears as a bright, clear line even in noisy scans.
- Diaphragm Detection: The system uses a PulseNet encoder paired with a Support Vector Machine (SVM) to identify frames containing the diaphragm. Temporal smoothing is applied to prevent "losing" the diaphragm in isolated noisy frames.
- Automated Initialization: The pipeline uses the Structural Similarity Index Measure (SSIM) to compare the first frame against a pre-labeled bank, selecting the best match to "seed" the segmentation.
- Segmentation and Propagation: An adaptive memory model propagates the initial mask across the video, exploiting temporal consistency.
- Liver Inference: The model identifies the endpoints of the segmented diaphragm and links them to the edges of the ultrasound's field of view, applying a geometric correction for the fan-shaped beam.
Key Results
The "anchor" approach proved significantly more robust than traditional direct segmentation. The PulseNet + SVM detection model achieved an F1 score of 0.96, outperforming standard ResNet-50 benchmarks.
As shown in the frame analysis above, the final predicted output successfully smooths over gaps in raw detection, ensuring a continuous identification of the diaphragm. Quantitatively, the pipeline achieved a mean IoU of 0.829, a substantial improvement over the 0.622 achieved by direct liver segmentation. Qualitatively, the pipeline was far more effective at capturing the full extent of the liver, including critical periportal areas.
Discussion and Future Work
By treating liver identification as a geometry problem anchored to a reliable landmark, this pipeline overcomes the limitations of low-quality POCUS data. However, the current model assumes the diaphragm is visible in the frame and was developed primarily on healthy adult data. Future work will focus on validating the pipeline in cases of advanced disease where liver tissue may be heavily distorted and integrating it into a field-ready diagnostic tool for automated schistosomiasis staging.
Summary updated