📄 infectious diseases

SchistoTrackNet: machine learning for diagnosis of schistosomiasis-associated periportal fibrosis from ultrasound images

This paper introduces SchistoTrackNet, the first deep learning model capable of automatically diagnosing varying severities of schistosomiasis-associated periportal fibrosis from ultrasound images, offering a scalable solution for large-scale surveillance in low-resource settings to support global elimination efforts.

Original authors: Ockenden, E. S., Anguajibi, V., Mpooya, S., Ntegeka, B., Mugume, T., Nabatte, B., Kabatereine, N. B., Noble, A., Chami, G. F.

Published 2026-06-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ockenden, E. S., Anguajibi, V., Mpooya, S., Ntegeka, B., Mugume, T., Nabatte, B., Kabatereine, N. B., Noble, A., Chami, G. F.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The Big Picture: A New "Smart Eye" for a Silent Killer

Imagine a parasite called a "blood fluke" that lives in the water of rural Africa. When people drink or swim in this water, the parasite gets inside them. Over time, these parasites lay eggs that get stuck in the liver, causing it to scar and harden. This scarring is called fibrosis.

For a long time, doctors have struggled to see how bad this scarring is. They use ultrasound machines (which take pictures of the inside of the body using sound waves), but reading these pictures is like trying to read a map in the dark. It requires years of training, and even then, two different doctors might look at the same picture and disagree on what they see.

This paper introduces SchistoTrackNet, a new computer program (an AI) that acts like a "super-reader" for these ultrasound pictures. Its job is to look at the pictures and automatically tell doctors exactly how severe the liver scarring is, from mild to severe.

The Problem: The "Map" Was Hard to Read

Doctors currently use a set of rules called the Niamey Protocol to diagnose liver scarring. Think of this protocol like a coloring book with specific patterns:

Pattern B: A bit of dust or faint lines (might not even be the parasite).
Pattern C: Rings or pipes around the blood vessels (mild scarring).
Pattern D: A thickening around the main pipe (moderate scarring).
Pattern E/F: Big, blocked pipes reaching the edge of the liver (severe scarring).

The problem is that these patterns look very similar. A doctor might look at a picture and think, "Is that a ring (C) or a thick pipe (D)?" Because the pictures are hard to interpret, doctors often need a second expert to double-check their work. But in poor, rural areas, there aren't enough experts to do this double-checking.

The Solution: Teaching a Computer to See

The researchers gathered 1,533 ultrasound pictures from people in rural Uganda. They used these pictures to train a computer brain (a deep learning model) to recognize the different patterns.

They tried five different types of computer brains:

SchistoTrackNet: This one was special. Before learning about liver scarring, it was already trained on pictures of baby fetuses (from a different dataset). It's like hiring a mechanic who already knows how engines work, then teaching them specifically about your car.
SchistoTrackNet-Random: Same design, but started with no prior knowledge (like a mechanic who just graduated school).
Three "Transformer" models: These are newer types of AI that look at the whole picture at once, rather than piece by piece.

The Results: The Computer vs. The Human

The researchers tested these computer brains on new pictures they had never seen before. Here is what happened:

The Winner: SchistoTrackNet was the best. It got the diagnosis right about 82% of the time.
The Comparison: The researchers also had a second human doctor look at the pictures after the first doctor took them (without knowing what the first doctor said).
- The two human doctors only agreed with each other about 54% of the time.
- The computer (SchistoTrackNet) agreed with the first doctor 77% of the time.

The Analogy: Imagine two people trying to identify different types of clouds in a photo. They often argue about whether it's a "cumulus" or a "stratus." The computer, however, looks at the photo and says, "That's definitely a cumulus," and it agrees with the first person much more often than the second person does.

What the Computer Got Wrong (and Right)

The computer was very good at spotting healthy livers and severe scarring. However, it sometimes got confused between patterns that look similar, specifically:

Pattern B (faint lines) vs. Pattern C2 (rings).
Pattern E/F (severe blockage) vs. Pattern C2.

The paper notes that this confusion happens because these patterns often appear in the same "angle" or view of the liver. It's like trying to tell the difference between a red ball and a red apple when they are both sitting in the same spot; the computer (and the human) gets confused by the angle, not just the object.

Why This Matters (According to the Paper)

The paper claims this is the first time a computer has been built to automatically distinguish between these specific types of liver scarring caused by this parasite.

It's a "Second Reader": The computer can act as a second pair of eyes to help confirm a diagnosis, which is crucial because human doctors often disagree.
It Handles Real Life: The computer worked well even when patients hadn't fasted (which makes gas in the stomach that blocks the view), had fatty livers, or were pregnant. It didn't get confused by these extra factors.
It's Not Perfect Yet: The paper admits the computer needs more testing on data from other countries and with cheaper ultrasound machines before it can be used in real clinics.

Summary

Think of SchistoTrackNet as a new, highly trained assistant for doctors in rural Africa. It looks at ultrasound pictures of livers and says, "This is mild scarring," or "This is severe." It makes fewer mistakes than a second human doctor trying to guess from a photo alone. While it still needs more practice, it offers a promising way to help doctors diagnose a dangerous disease that is currently hard to spot.

Technical Summary of SchistoTrackNet

Problem Statement
Liver fibrosis is a leading cause of mortality and disability in low- and middle-income countries (LMICs), with intestinal schistosomiasis (specifically Schistosoma mansoni) being a historically underestimated cause in sub-Saharan Africa. While the World Health Organization (WHO) aims to eliminate schistosomiasis as a public health problem by 2030, current guidelines rely on infection intensity as a proxy for morbidity, despite evidence showing no correlation between current infection levels and liver fibrosis. The diagnosis of schistosomiasis-associated periportal fibrosis (PPF) relies on the Niamey protocol, which uses B-mode ultrasound imaging. However, this protocol requires extensive expertise and exposure to diverse fibrosis patterns, resources often lacking in endemic regions. Furthermore, there is a critical gap in automated, clinically validated tools for the differential diagnosis of PPF severity. Existing deep learning applications in this domain are limited, with few studies utilizing second readers to validate diagnoses, and no known models capable of distinguishing the distinct, varying severity patterns of schistosomal PPF.

Methodology
The study utilized data from the SchistoTrack cohort study, involving 1,533 participants (aged 5–84) from three rural districts in Uganda (Pakwach, Buliisa, and Mayuge). Ultrasound images were acquired using Philips Lumify systems with C5-2 curvilinear transducers. The dataset comprised 3,710 curated images of liver fibrosis patterns and 400 frames from healthy participants (derived from video sweeps of 100 healthy individuals, as static images were not captured for healthy livers).

The researchers developed and evaluated five deep learning models for a 6-class classification problem (Healthy, B, C1, C2, D, and E/F):

SchistoTrackNet: A Convolutional Neural Network (CNN) using a PULSENet encoder (pre-trained on fetal ultrasound anatomy) initialized with supervised contrastive (SupCon) loss.
SchistoTrackNet-Random: A CNN with the same architecture as SchistoTrackNet but with random weight initialization.
ViT-SupCon, ViT-Focal, ViT-CE: Three Vision Transformer (ViT) models initialized with ImageNet weights, trained using SupCon loss, Focal loss, and Cross-Entropy loss, respectively.

The models were trained using 10-fold cross-validation (stratified by participant) and evaluated on a held-out test set. Performance was assessed using accuracy, balanced accuracy, F1-score, sensitivity, specificity, and Cohen's $\kappa$ . The models were compared against a "second sonographer" who re-read images in a blinded setting (without patient context) to simulate a validation scenario. Additionally, GRAD-CAM and attention rollout techniques were used to visualize model activation and interpretability.

Key Results

Model Performance: SchistoTrackNet achieved the highest performance among all models on the test set, with an accuracy of 0.8222, balanced accuracy of 0.8121, and F1-score of 0.8223. Statistical analysis (Friedman's $\chi^2$ -test) indicated significant differences in performance between models, with SchistoTrackNet outperforming ViT-Focal and ViT-CE in F1-score.
Comparison with Human Readers: SchistoTrackNet demonstrated higher agreement with the field sonographer (Cohen's $\kappa$ = 0.772) than the second sonographer did with the field sonographer ( $\kappa$ = 0.54). The model classified more images correctly than the second sonographer agreed upon for most patterns.
Pattern Specificity: The model showed high sensitivity for the D pattern (91.4%) and high Positive Predictive Values (PPV) for B, C1, and D patterns. However, confusion remained between patterns sharing similar transverse views (B, C2, and E/F), mirroring the confusion observed in human readers.
Robustness: Patient-level factors such as age, sex, fasting status, pregnancy, and comorbidities (hepatitis B, fatty liver) had little to no significant effect on model error rates, suggesting the models are robust to common clinical variations.
Interpretability: Activation maps (GRAD-CAM) indicated that the model focused on relevant anatomical features, such as the main portal vein for D patterns, though some activation occurred in areas of gas interference or the diaphragm.

Significance and Claims
The authors claim to present the first deep learning-based models capable of fully automated, differential classification of schistosomal periportal fibrosis patterns. The study demonstrates that these models can perform comparably to, and in some metrics better than, expert human sonographers acting as second readers, despite lacking the contextual information (patient history, probe position) available during the initial scan.

The paper posits that these tools have the potential to enable automated reading of ultrasound images, thereby facilitating large-scale surveillance of schistosomiasis morbidity. This capability is framed as a necessary step toward meeting WHO targets for eliminating schistosomiasis as a public health problem by 2030. The authors suggest that the models could serve as a starting point for tools to support capacity building in resource-limited settings and contribute to updating the Niamey protocol by providing standardized image banks and training aids. The study acknowledges limitations, including the use of video frames for the healthy class, data from a single country and ultrasound machine type, and the need for further clinical trials to assess impact on patient outcomes.