Human-computer collaboration for skin cancer recognition

A multi-part reader study rather than a single experiment. Five sub-studies share a CNN and a web platform but differ in participants, images and design. The headline sub-study is the interaction study; the coded fields describe that study unless noted.

AI system. A ResNet34 pretrained on ImageNet and fine-tuned on the HAM10000 dataset to classify seven diagnostic categories: melanoma, basal cell carcinoma, actinic keratosis and intraepithelial carcinoma, melanocytic nevus, benign keratinocytic lesion, dermatofibroma and vascular lesion. Class-weighted cross-entropy loss, Adam, learning rate initialised at 0.0001, batches of 32, images cropped and resized to 224 by 224, maximum 100 epochs with early stopping. Fivefold cross-validation; the single best network on the held-out validation set was used with raters. Multiclass probabilities were produced by softmax. CBIR was implemented on the same CNN by extracting the target image's feature vector and retrieving the four closest images per class by cosine similarity from HAM10000.

Interaction study. N = 302 raters from 41 countries, recruited through International Society of Dermoscopy mailing lists and social media, comprising 169 board-certified dermatologists (56.0%), 77 dermatology residents (25.5%) and 38 general practitioners (12.6%). Raters registered, self-reported years of dermatoscopy experience in five bands from less than 1 year to more than 10 years, and had to pass screening tests of domain-specific tasks so that self-reported experience could be checked against actual skill. Images came from the ISIC 2018 benchmark test set: 1,412 of 1,511 images were used, drawn from Vienna, Queensland, Turkey, New Zealand, Sweden and Argentina to give diversity of skin types, with histopathology, sequential imaging biology, expert consensus or confocal imaging as ground truth. Batches of 28 images were generated, each containing four randomly selected examples of every class. Protocol: each rater diagnosed the batch first without any support, then again with one of the four support types, so the unassisted judgement is recorded before the AI is revealed. A stratified randomisation balanced the four support types across disease categories. Only the first five tests per rater were included, to avoid biasing results toward high-repetition raters, and tests with fewer correct answers than chance were excluded. The study ran from 29 May 2019 to 15 January 2020 and yielded 512 tests and 13,428 ratings.

Extended interaction study, faulty AI. N = 155 raters, 15 January to 18 February 2020, multiclass probabilities only. Each test contained one image per diagnosis with unaltered probabilities, two images whose probabilities had been shuffled so that the top class favoured a random incorrect diagnosis, and eight images from the telemedicine set. Image sources were not disclosed to raters.

Telemedicine sub-study. Images were reused from the intervention arm of a randomised trial on mobile teledermoscopic self-examination: 93 of 98 participants, 1,521 self-made photographs of 596 lesions. Ground truth was face-to-face examination by a board-certified dermatologist or histopathology. Triage performance was computed on the CNN alone at a malignancy cutoff of 0.17, chosen because that was the cutoff raters had implicitly adopted in the interaction study. The telemedicine support analysis used the 155 extended-study raters.

Second-opinion sub-study. 11 raters, 8 board-certified dermatologists and 3 dermatology residents, each of whom had diagnosed and managed at least two suspicious lesions face to face between April and September 2019. 79 cases with definite histopathologic diagnoses. Physicians were asked to reconsider diagnosis and management with AI-based multiclass probabilities, without being told they had previously managed the patient or what they had decided. Management options were no intervention, monitor or excise. No time constraints were set.

Knowledge transfer sub-study. 189 fourth-year medical students, of 650 invited and 200 who agreed, who answered more than 50% of test questions. Students received a 30-minute introductory dermoscopy lecture, diagnosed 25 stratified test images, then were taught one additional clue derived from Grad-CAM analysis, that sun-damaged surrounding skin marks pigmented actinic keratoses, and repeated the test. No AI support was present at decision time in this sub-study.

Measures and analysis. Diagnostic accuracy as the proportion of correct specific diagnoses; rate of changing the initial diagnosis; time to answer, normalised within rater and modality and used as a surrogate marker for confidence, with answers faster than the individual mean treated as confident; number of CBIR thumbnails enlarged as an engagement measure. Paired and unpaired t-tests, Wilcoxon signed-rank tests and chi-squared tests as appropriate, all p values Holm-Bonferroni corrected, two-sided, alpha 0.05, analysed in R 3.6.2.

Only one of the four representations of the same CNN's output produced any benefit, which is the paper's most transferable result: the format in which AI output is presented determines whether it helps at all, independently of how accurate the model is.

AI-based multiclass probabilities raised rater accuracy from 63.6% to 77.0%, an increase of 13.3% (95% CI 11.5% to 15.2%, p = 4.9 x 10-35, two-sided paired t-test, t = 14.5, d.f. = 301, n = 302). Neither the AI-based probability of malignancy nor the CBIR representation produced any improvement, despite being derived from the identical network. The authors read this as a match between the form of support and the structure of the task: a binary malignancy probability suits a binary management decision but not a seven-way diagnostic one, and CBIR requires the rater to extrapolate a diagnosis from visual similarity, which is neither simple nor concrete. The behavioural data support the cost interpretation. Raters took 16.5 s (95% CI 14.5 to 18.6) to engage with CBIR against 4.6 s (95% CI 4.3 to 4.9) for multiclass probabilities, 5.2 s for malignancy probability, 4.5 s for crowd probabilities and 5.6 s with no support at all, and the number of CBIR thumbnails they opened declined as the study went on. Raters progressively disengaged from the support that cost the most and returned the least.

Benefit was inversely related to experience. The correlation between net gain and rater experience was r = -0.18 (95% CI -0.28 to -0.07, p = 1.5 x 10-2, n = 302). The least experienced raters changed their initial diagnosis on 26.0% of cases (95% CI 21.3% to 30.7%) against 14.7% for experts (95% CI 9.9% to 19.6%). Experts gained only when they were unsure: net gain 13.4% (95% CI 6.3% to 20.6%) when not confident, against -0.7% (95% CI -6.8% to 5.4%) when confident. The authors extract a usable heuristic from this, that a confident expert should not consult the system, and note that confident experts were usually right. Confidence governed switching across the whole sample: raters changed their diagnosis on 14.7% of confident cases (95% CI 12.6% to 16.8%) against 37.5% of non-confident cases (95% CI 34.0% to 41.0%, p = 1.9 x 10-25). The least experienced raters, by contrast, accepted contradicting AI advice even when confident.

Faulty AI reversed the gain for every group, experts included. When the top-class probabilities were shuffled to favour a random incorrect diagnosis, a median accuracy gain of 9.5% (p = 1.2 x 10-12, two-sided paired Wilcoxon signed-rank, n = 155) became a median loss of 6.3% (p = 6.0 x 10-13, n = 155). The authors put the mechanism plainly: the trust that makes benefit possible is the same trust that makes harm possible, and they leave open whether interpretability techniques would mitigate it.

Switching behaviour was governed by the shape of the probability distribution rather than by its content. Changes occurred almost exclusively toward the AI's top-ranked class; raters held a disagreeing decision only when the AI ranked it second or third. They switched more often when the gap between the AI's probability for their own initial choice and for the AI top class was large. Big, clear winners moved raters; low-probability classes barely registered.

The gain was also unevenly distributed across disease categories, concentrated on pigmented actinic keratoses and intraepithelial carcinoma, where it reached 31.5% (95% CI 22.9% to 40.1%, n = 43 images). This asymmetry is what generates the paper's final result.

The two workflow simulations both worked. In telemedicine triage, the CNN's specific-diagnosis accuracy dropped sharply on patient-taken images, from 76.2% to 53.9% (p = 8.9 x 10-14, chi-squared, n = 1,430), yet it still identified 95.2% of patients with skin cancer at 59.2% specificity, because multiple images per patient compensate for weak per-image sensitivity. In the second-opinion simulation, dermatologists reviewing their own prior face-to-face cases without knowing it switched from excise to monitor on 15.5% of benign lesions, 7 of 45, with no compensating movement in the wrong direction on malignant lesions, and specific diagnostic accuracy rose from 55.6% to 75.0% (p = 0.029, n = 11 raters). Aggregation also helped: bootstrapped human collectives reached 73.7% against 64.8% for single raters, and combining collectives with AI multiclass probabilities reached 81.0% (95% CI 78.2% to 83.9%, p = 8.6 x 10-9), above the CNN alone at 76.9%.

Finally, an XAI artefact was converted into human learning. Grad-CAM showed the network attending to the non-lesion background more for actinic keratoses than for other classes (mean ratio 0.48 against 0.32, p = 4.6 x 10-12), which the authors interpret not as a Clever Hans artefact but as a real clue: chronic sun damage surrounds actinic keratoses. Teaching that single clue to medical students raised correct diagnosis of pigmented actinic keratoses from 32.5% (95% CI 30.0% to 35.0%) to 47.3% (95% CI 43.9% to 50.8%, p = 3.6 x 10-13) and overall accuracy from 55.2% to 59.1% (mean difference 3.7%, 95% CI 2.4% to 5.3%, p = 3.4 x 10-6, n = 189). The transfer runs from model to human and persists without the model present.