Did AI heatmaps improve glioblastoma survival judgements?

The answer: assistance helped in a small reader study, while the audit exposed a serious weakness

The clearest result is not that a coloured heatmap made a brain-tumour model trustworthy. Four clinicians were more accurate when they could see the model's survival prediction and were more accurate again when heatmaps were added. Mean area under the receiver-operating-characteristic curve rose from 0.718 with one MRI slice alone to 0.760 with the numerical AI output and 0.781 with AI plus explanations. Compared with the unassisted phase, the odds of a correct judgement were higher with the score and heatmaps, at an odds ratio of 1.598 and p below 0.01.

But the heatmaps' additional accuracy benefit over the AI score alone did not meet the study's significance threshold: p was 0.0808. The more consequential finding came from auditing the model itself. Its survival predictions depended strongly on a narrow band of unusually bright pixels around the skull and contrast-enhancing tumour. Removing only the brightest 5% of foreground pixels reduced AUC from 0.89 to 0.642. That dependence may reflect useful post-operative anatomy, a scanner-specific artefact or both. Until the model is tested on independent hospitals and full clinical imaging, the paper is evidence for careful audit—not evidence for deployment.[1]

What the model was asked to predict

The researchers audited a previously developed convolutional neural network that classified overall survival after glioblastoma treatment as short term, meaning 332 days or less, or long term, meaning more than 332 days. The threshold was the median survival in the cohort. Input consisted of post-operative, pre-radiotherapy contrast-enhanced T1 MRI. Nine central two-dimensional slices were extracted for each patient, although the clinician experiment used one central slice per patient so the readers and model saw the same image.

The dataset contained 175 patients and 1,575 slices. Patients—not individual slices—were divided into training, validation and test groups to reduce leakage: 104 patients supplied 936 training slices, 34 supplied 306 validation slices and 37 supplied 333 held-out test slices. The test group contained 18 short-term and 19 long-term survivors. This partitioning is stronger than randomly mixing slices from the same person across development and evaluation, but every patient still came from one French cancer centre and one imaging sequence.[1]

Why the audit used more than a single saliency map

A heatmap can show which pixels are associated with a prediction without proving that those pixels caused it. The team therefore combined three types of explanation. Grad-CAM, layer-wise relevance propagation and LIME mapped where the network appeared to look. Linear probes tested where survival classes became separable inside the network. Targeted ablation then removed candidate image regions or the brightest pixels and measured what happened to predictive performance. Agreement across localisation, internal representation and intervention provided stronger evidence of a shortcut than any one attractive visualisation could.

The three saliency methods did not behave identically. No method ranked first across all five evaluation criteria, and Grad-CAM was unstable under some noise tests. LIME relevance correlated strongly with pixel intensity, with mean Spearman correlation of 0.69, while Grad-CAM's correlation was 0.28. About 76% of the high-intensity signal lay in the skull and about 14% in contrast-enhancing tumour. Masking the whole tumour reduced AUC from 0.89 to 0.83; masking the brightest 1% dropped it to 0.79, and masking the brightest 5% dropped it to 0.642. A model can therefore post an impressive summary score while relying on a fragile visual cue.[1]

How the clinician experiment worked

Three radiation oncologists and one radiologist reviewed all 37 held-out cases. They were the available clinicians at the participating centre, ranging from a doctor in training to experienced consultants. Each clinician first judged survival from the MRI slice alone, then saw the model's class and confidence score, and finally saw the score with fine-grained and regional heatmaps. That created 37 patients multiplied by four readers multiplied by three phases, or 444 observations. The investigators used clustered bootstrapping and generalised estimating equations to account for repeated judgements of the same cases.

The design makes within-reader change visible, but it cannot isolate every cause of that change. The phases were cumulative and always occurred in the same order, so readers had already seen each case before reaching the heatmap phase. Case order was changed between phases, but the study could not rule out memory or practice effects. All readings happened in one session, no scheduled break was included, and clinicians used static PDF images on office monitors rather than complete three-dimensional studies with adjustable windowing. Reading time was not recorded.[1]

What changed when clinicians saw the explanations

The standalone model had an AUC of 0.869 on the 37 central slices. Clinician AUC improved stepwise across the three phases, but the study does not establish that clinicians plus AI were superior to the model alone. Adding the AI prediction increased the odds of a correct response relative to the baseline phase, with an odds ratio of 1.336 and p below 0.05. Adding heatmaps produced the larger comparison with baseline, but not a statistically significant improvement over the score-only phase.

The heatmaps nevertheless appeared to change how readers used the model. An interaction test between phase and the AI score was significant at p below 0.01. False positive judgements—patients incorrectly classified as likely to survive longer—fell from 19 with the AI score to 13 when heatmaps were added, while false negatives changed from 21 to 22. Average confidence did not rise significantly between baseline and the final phase, although confidence remained strongly anchored to the AI score. This is a behavioural signal from a tightly controlled experiment, not proof that treatment decisions or patient outcomes would improve.[1]

What this means for patients and clinical teams

For a person with glioblastoma, a survival category can influence expectations and conversations about treatment. A model that confuses a local imaging artefact with prognosis could create false reassurance or undue pessimism. The study's binary 332-day threshold also compresses a complex outcome into two groups. It does not estimate an individual's survival time, account for every treatment and molecular factor, or show whether a clinician would change radiotherapy, surgery, chemotherapy or supportive care appropriately.

For radiologists, oncologists and medical-physics teams, the paper offers a practical audit lesson. Do not treat a plausible heatmap as validation. Test whether removing the highlighted feature changes the prediction, inspect what the network encodes at different depths and repeat the analysis on independent scanners. Explanations should help a clinician challenge a model, not simply make its score easier to accept. Procurement and governance should therefore request external validation, scanner and protocol sensitivity, calibration, failure examples and the exact way explanations were tested.

The work was funded by Région Normandie and an innovation grant from Amgen France. The authors declared no competing interests. Funding does not invalidate the results, but it belongs in the evidence record—particularly when a model concerns cancer treatment and the clinical study has a small denominator.[1]

What would change the assessment

The next essential test is locked external validation using data from multiple hospitals, MRI manufacturers, field strengths and acquisition protocols. The same ablation and probing audit should be repeated before any harmonisation and after any proposed intensity normalisation. If the bright-pixel dependence reflects a reproducible post-operative anatomical feature rather than one centre's image characteristics, that should remain measurable across sites. If performance collapses, the model should be redesigned rather than cosmetically explained.

Clinical evidence needs a larger, counterbalanced multi-reader study using full volumetric imaging and the clinical information specialists normally receive. Readers should be randomised to assistance conditions or use a design that limits recall, and investigators should record decision time, disagreements, calibration and downstream management choices. A prospective study would then need to show that assistance improves appropriate care without increasing harmful overconfidence, delay or inequality. Until those steps are complete, the defensible conclusion is narrow: multi-method explainability exposed a shortcut and changed four clinicians' use of an AI score, but it did not establish clinical benefit.[1]