SchistoTrackNet: a neural network reads liver ultrasound to spot the fibrosis of bilharzia (medRxiv, 2026)

Posted on 2 June 2026 on medRxiv, this preprint from a team at Oxford and the Ugandan Ministry of Health trains a neural network, SchistoTrackNet, to read liver ultrasounds and spot the periportal fibrosis caused by schistosomiasis, a neglected parasitic disease. On 3,710 images from the rural SchistoTrack cohort in Uganda, the model correctly classifies six fibrosis patterns 82.2% of the time and agrees more closely with the sonographer who took the scan (Cohen's kappa 0.77) than a second sonographer re-reading the same image blind (0.54). The topic matters, the method is careful, and separating patients between training and test avoids a classic mistake. But the ground truth remains the subjective opinion of a single human reader — with no histological confirmation —, the data come from a single country and a single type of ultrasound machine, and the most advanced, most dangerous fibrosis is caught only half the time.

The context

Schistosomiasis (or bilharzia) is an infection caused by a freshwater parasitic worm; more than 700 million people are at risk in sub-Saharan Africa. In its intestinal form, due to the parasite Schistosoma mansoni, eggs laid in the abdominal veins are carried to the liver, where the chronic immune reaction can drive periportal fibrosis: a scar-like thickening of the tissue around the liver's vessels, which disrupts its circulation and can eventually cause fatal digestive bleeding. Long underestimated, this fibrosis is a major cause of liver morbidity on the continent.

The reference diagnostic tool is portable ultrasound (POCUS), cheap and carriable down to the village. The reading standard is the Niamey protocol, published by the WHO in 2000: it describes a series of image patterns, lettered A to F, from the first fibrous streaks (B patterns) to the obstructed vessels of severe stages (E and F patterns). The trouble, the authors note, is that this protocol "requires extensive expertise and exposure to many cases," whereas the hardest-hit countries lack precisely the staff trained in sonography. Hence the idea of a software assistant able to recognize these patterns — still almost untouched ground: only two prior studies exist, one of them trained on barely 160 images.

The method

The work, authored by Eloise Ockenden, Goylette F. Chami (Nuffield Department of Population Health, Oxford) and co-authors from the Ugandan Ministry of Health, builds on the SchistoTrack cohort: a prospective field study run in three rural districts of Uganda (Pakwach, Buliisa and Mayuge). Images were acquired by four experienced sonographers on tablets fitted with Philips probes. The dataset gathers 3,710 fibrosis-pattern images from 1,433 participants, supplemented by 400 images of fibrosis-free livers (frames extracted from videos in 100 healthy participants) to form a "normal liver" class.

The ground truth — the label the model must learn to reproduce — is the pattern assigned by the sonographer at the time of the exam, following the Niamey protocol. The eight original patterns were grouped into six classes: B (early fibrosis, not necessarily from schistosomiasis), C1, C2, D, the E/F pair (the most severe) and the healthy liver. So this is a multiclass classification task of the degree of fibrosis, not a simple "sick / not sick."

On the model side, the authors compare five networks. SchistoTrackNet is a convolutional network (a CNN, a family specialized in image analysis) with a VGG-16 architecture, initialized not at random but with the weights of PULSENet, an encoder pre-trained to recognize fetal ultrasounds — the idea being to start from a network that "already knows" the very particular texture of ultrasound before specializing it on the liver. It is trained by supervised contrastive learning, a technique that teaches the network to pull together images of the same pattern and push apart those of different patterns. The competitors are three Vision Transformers (ViT), a more recent architecture that cuts the image into patches and relates them. The train/test split is 90/10 with ten-fold cross-validation and — crucially — it is done per participant: no image from a given patient appears in both training and test, which shuts the door on data leakage.

The results

On the test set, SchistoTrackNet achieves the best accuracy, 82.2%, a balanced accuracy of 81.2% and an F1 score of 0.82 (F1 is the harmonic mean of precision and recall). Its agreement with the field sonographer, measured by Cohen's kappa — an index that equals 1 for perfect agreement and 0 for chance-level agreement —, reaches 0.77. Specificity exceeds 90% for every pattern but one (89.7% for C2), and sensitivity ranges from around 86% for the intermediate patterns up to 97.5% for the healthy liver.

The most telling result is the comparison with a second human sonographer, re-reading the same images blind a few months later. Agreement between two human experts was only 0.54 — and dropped to 46% for the faintest B patterns. In other words, on this task the model sticks more faithfully to the original diagnosis than another specialist does: a strong argument for use as an automatic "second reader."

Clinical translation: these flattering numbers hide a sore point. On the E/F pattern, the most advanced stage — the one that truly threatens survival, associated with esophageal varices —, the model's sensitivity is only 57.9%. Concretely, out of 100 patients with the most severe fibrosis, the model would miss about 42. For the intermediate stages (D pattern), sensitivity exceeds 90% with a positive predictive value of 0.89, which is solid. So the model is good at sorting the bulk of the flow, but least reliable precisely where an error costs the most.

What is good

A neglected disease, taken seriously. Schistosomiasis affects hundreds of millions of people among the poorest, and draws only a tiny fraction of medical-AI research. Building and using what the authors present as the largest imaging dataset on schistosomal morbidity in sub-Saharan Africa, from a real field cohort rather than lab images, is in itself a useful contribution.

A leak-proof data split. Splitting images by participant, not at random, ensures that no patient "leaks" from training into test. It is the most basic methodological hygiene, but it is so often neglected in the AI-health literature — where spectacular AUCs collapse the moment this point is controlled — that respecting it deserves credit.

Honesty about human subjectivity. Rather than presenting the sonographer as absolute truth, the authors measure the agreement between two experts (kappa 0.54) and display it. This reframes the problem: it is not about matching a perfect gold standard, but about providing a reproducible reading where even specialists diverge. The cohort is also inclusive — no exclusion of fatty livers or scans degraded by intestinal gas —, which makes the evaluation more realistic.

What is less good

A fragile ground truth (failure mode: the biased comparator). The model learns to reproduce the judgment of a single sonographer, with no confirmation by biopsy or any other independent test. Yet that human judgment is itself uncertain, as the inter-reader kappa of 0.54 shows. "Agreeing with the first reader better than a second reader does" may therefore just as well mean that the model has learned the habits and biases of the original operator, not the biology of the fibrosis.

A whiff of shortcut, and a "healthy" class that is too easy (failure modes: shortcut learning and misleading metric). The model's confusions concentrate among patterns sharing the same ultrasound view (B, C2, E/F), which, the authors admit, "may be evidence of the model learning the transverse view of the liver rather than the severity" of the disease — a textbook case of shortcut learning, where the network latches onto a spurious cue. Grad-CAM activation maps confirm the risk: the model sometimes lights up on the diaphragm or on gas interference, not on the lesion. Moreover, the "healthy liver" class is made of video frames of a different nature from the rest; the authors acknowledge this makes the distinction artificially easy and probably inflates the overall accuracy.

No external validation, and partial reproducibility (failure modes: population bias and reproducibility). All performance is measured on a single country, with a single type of ultrasound machine (Philips) and the same operators. Nothing guarantees the model holds up on a cheaper portable device or in another setting — the authors acknowledge this and make it a priority for future work. Code is provided as supplementary material, but the images cannot be shared (data protection), no model weights are released, and some predictive values, computed without a fixed sensitivity threshold, are "possibly inflated" by the authors' own admission. Finally, this is a retrospective analysis of already-collected images, with no prospective clinical trial.

What it changes

For the research community, the paper sets a marker: it establishes a serious dataset and a baseline for a disease nearly absent from the medical-AI landscape, and shows that an encoder pre-trained on fetal ultrasound can be repurposed toward the liver. The path to pursue is clearly laid out — validation on other countries and other machines, and above all anchoring the ground truth on something other than a single reader.

For clinicians, nothing changes in practice today: this is a preprint not yet peer-reviewed, with no external validation or prospective trial. The idea of a digital "second reader," able to bring consistency where two experts agree only halfway, is nonetheless credible and worth following — provided one keeps in mind that the model is weakest on the most severe forms, exactly where it is needed most.

For patients and the public, the lesson is nuanced. Putting AI to work on a disease of poverty, on a device as simple as a pocket ultrasound, is exactly the kind of application with high potential impact for under-resourced health systems. But "82% accuracy" does not mean "ready to use": between a promising demonstration on a single-country cohort and a tool deployable in routine, the decisive steps of real-world validation remain.

Going further

The full preprint is available on medRxiv (DOI 10.64898/2026.06.01.26354609), posted 2 June 2026 under a CC BY 4.0 license, by a team from the Nuffield Department of Population Health (University of Oxford) and the Vector-Borne and Neglected Tropical Diseases Control Division of the Ugandan Ministry of Health. For the ultrasound reading framework of fibrosis, the Niamey protocol (WHO, 2000) remains the reference; for the disease itself, the World Health Organization's fact sheets on schistosomiasis offer a reliable entry point.