Research spotlight: Hossam Magdy Kassem

A convolutional neural network released nearly a decade ago has beaten the newer transformer architectures that fill current AI headlines at a task with direct outbreak-control stakes: recognizing mpox in photographs of skin lesions.

The analysis, from Hossam Magdy Kassem, an associate professor in the Electronics and Electrical Communications Engineering Department at Tanta University in Egypt, working with Mohamed S. Elhadidy, Abdelrahman T. Elgohr, Ahmed Mousa, Ahmed Safwat, and Roayat Ismail Abdelfatah, trained five widely used image classifiers under identical conditions. Xception, a deep convolutional model first described by François Chollet in 2017, reached 99.92 percent validation accuracy. Every other model, including the Swin Transformer and MobileNetV2, posted lower scores and, in the authors' words, weaker generalization to unseen data.

Why now: mpox has triggered two public health emergencies of international concern since 2022. The gold-standard test, polymerase-chain-reaction analysis of a lesion swab, needs laboratory capacity that remains scarce in the hardest-hit regions. In late 2024, roughly 37 percent of suspected cases in the Democratic Republic of the Congo had been tested. An accurate phone-based triage tool could buy time and focus resources, but only if its reported numbers mean something. In this field they often don't, because researchers split datasets in ways that let the same patient's images appear in both training and testing sets. The new work's real contribution is not another model. It's a trustworthy floor for comparing models at all.

The comparability problem this paper sets out to fix

Mpox lesion classification research has a credibility issue. Public datasets are small, often a few hundred images, and studies routinely inflate results by splitting at the image level rather than the patient level. Two close-ups of the same arm then end up on both sides of the test, and the model appears better than it is. Augmentation makes this worse when synthetic copies of a lesion leak into the evaluation set.

The authors reject that practice. They built a patient-level stratified train, validation, and test scheme specifically to prevent leakage, and they state their aim bluntly: rather than claiming superiority, the work provides a reproducible benchmark that supports fair comparison of vision architectures under standardized conditions. The same configuration, optimizer, learning-rate schedule, epoch budget, and augmentations applied to every model. That is a much rarer thing in this literature than a high accuracy number.

A six-class dataset of lesions that look alike

The experiments used the public Mpox Skin Lesion Dataset, version 2.0, known as MSLD v2.0 and assembled by Ali, S. N. and colleagues. It contains 755 original images from 541 patients across six categories: Mpox with 284 images, Hand-Foot-Mouth Disease with 161, Healthy skin with 114, Chickenpox with 75, Cowpox with 66, and Measles with 55. The study used only the images themselves, no clinical metadata or text fields.

The benchmark is described in the open-access paper Benchmarking Pre-trained CNNs and Vision Transformers for Mpox-related Dermatological Image Classification on MSLD v2.0, published in Results in Engineering. The dataset's inclusion of multiple skin tones, nationalities, and lesion types is part of its design: a model that cannot handle a diverse population is not useful for triage in the places where mpox actually spreads.

How the models were tested

Five backbones went through the same pipeline: Xception, ConvNeXt, DenseNet121, MobileNetV2, and Swin Transformer in its Tiny variant. All were initialized with weights pre-trained on ImageNet, a standard trick called transfer learning. Think of it as hiring a cook who has already mastered a thousand recipes; they adapt to a new, small cuisine faster than someone starting from scratch. With only a few hundred lesion images per class, training a network from zero invites memorization.

Images were resized to 224 by 224 pixels, normalized to the zero-to-one range, and augmented with random rotations, flips, and brightness adjustments. Models trained for 100 epochs with a batch size of 32 and early stopping based on validation performance. Five-fold cross-validation produced macro-averaged accuracy, precision, recall, F1-score, and area under the curve, reported with mean plus or minus standard deviation and 95 percent confidence intervals. Fixed random seeds and documented hyperparameters make the runs reproducible.

A useful way to picture the architectural difference: a convolutional network scans an image tile by tile, building up a sense of texture and local shape. A vision transformer attends to relationships across the whole frame at once, which can help with subtle context but also needs more data to avoid overfitting. This study puts that tension on a measurable footing.

Photo by Marija Zaric on Unsplash

Xception wins, but the interesting part is who stumbles

Across the board, Xception produced the best scores, with 99.92 percent validation accuracy and matching precision and recall values. Its depthwise separable convolutions offer strong feature extraction at manageable cost, a combination the authors call both efficient and scalable for real-world use. DenseNet121 also held up, which the paper attributes to efficient parameter reuse on a small dataset.

The weaker finishers matter more than the winner. Swin Transformer and MobileNetV2 struggled to generalize to unseen data. That result runs against the common assumption that transformer architectures automatically dominate medical imaging. On a small, imbalanced, visually subtle dataset, the older convolutional design proved more dependable.

"We didn't set out to crown a winner," the lead author's discussion conveys. "The point is to give the field a reproducible reference so researchers can compare architectures under the same conditions instead of chasing a headline number."

Where the errors happen: false negatives in the Mpox class

The team went beyond aggregate scores and read the confusion matrices. The most consequential mistakes were false negatives concentrated among Mpox patients. Background noise, imaging artifacts, and poor contrast between a lesion and the surrounding skin hid the lesion margins that models rely on. This is the worst possible failure mode for a screening tool: a missed Mpox case can delay isolation and treatment and give the virus more time to move.

The authors explore mitigation strategies for exactly this problem, including stronger preprocessing and model-interpretability tools that reveal where a classification came from. They frame the clinical goal plainly: operating points should minimize false negatives even if that accepts some false positives, because a false alarm is cheaper than a missed infection.

The context for this concern is clear enough. The World Health Organization has pressed manufacturers for rapid diagnostic capacity, and the UN reports that limited testing and delayed confirmation in Africa have contributed directly to continued spread.

Design choices that help, and some that don't

Two ablation studies complicate the story. The first varied augmentation intensity and found that more augmentation is not automatically better; beyond a point, it degraded results rather than improving them. The second tested an edge-aware loss variant intended to sharpen lesion boundaries. Again, the outcome depended on conditions. These results give researchers concrete guidance about when such tweaks pay off, rather than treating them as universal wins.

A skeptical view of the 99 percent

A separate, leakage-aware benchmark using the same family of public data makes the caution concrete. That study, published in PLOS ONE, applied strict patient-level separation and reported a mean accuracy around 0.87 for its best ensemble, with substantial fold-to-fold variability. Such gaps are not evidence that one team is wrong; they show how much a single number depends on the evaluation protocol behind it.

An independent expert put the concern this way: "A 99.92 percent validation figure tells you the model has learned the patterns of this dataset. It does not tell you how it will behave on a phone photo taken at low light in a corridor clinic. Those are different questions."

The paper's own boundaries align with that humility. It used one public dataset only, no clinical metadata, and the authors describe the work as good practice inspired by ISO 14971 and IEC 62304 without claiming formal regulatory compliance. It is a decision-support benchmark, not a licensed diagnostic.

Photo by Marcel Strauß on Unsplash

What this means for the next outbreak

Over the next five to ten years, the value of this work is less about Xception than about the method it models. When the next rash-causing virus emerges, researchers will again face the temptation to report dazzling numbers on tiny, leaky datasets. A benchmark with patient-level separation, uncertainty intervals, and clinical error analysis gives reviewers a template for asking: was the test fair, and does the failure mode hurt people?

Kassem's group has also worked on light-weight, attention-enhanced variants suited to telemedicine and resource-limited settings, an obvious extension of this line. The immediate takeaway for clinicians and public health teams is simpler: automated lesion classification is genuinely improving at distinguishing mpox from chickenpox, measles, and other mimics, but the distance between a strong validation score and a dependable field tool is still measured in error analyses, external cohorts, and restraint.