Prompt-driven Latent Domain Generalization for Medical Image Classification.

Deep learning models for medical image analysis easily suffer from distribution shifts caused by dataset artifact bias, camera variations, differences in the imaging station, etc., leading to unreliable diagnoses in real-world clinical settings. Domain generalization (DG) methods, which aim to train models on multiple domains to perform well on unseen domains, offer a promising direction to solve the problem. However, existing DG methods assume domain labels of each image are available and accurate, which is typically feasible for only a limited number of medical datasets. To address these challenges, we propose a unified DG framework for medical image classification without relying on domain labels, called Prompt-driven Latent Domain Generalization (PLDG). PLDG consists of unsupervised domain discovery and prompt learning. This framework first discovers pseudo domain labels by clustering the bias-associated style features, then leverages collaborative domain prompts to guide a Vision Transformer to learn knowledge from discovered diverse domains. To facilitate cross-domain knowledge learning between different prompts, we introduce a domain prompt generator that enables knowledge sharing between domain prompts and a shared prompt. A domain mixup strategy is additionally employed for more flexible decision margins and mitigates the risk of incorrect domain assignments. Extensive experiments on three medical image classification tasks and one debiasing task demonstrate that our method can achieve comparable or even superior performance than conventional DG algorithms without relying on domain labels. Our code is publicly available at https://github.com/SiyuanYan1/PLDG/tree/main.

The SLICE-3D dataset: 400,000 skin lesion image crops extracted from 3D TBP for skin cancer detection.

AI image classification algorithms have shown promising results when applied to skin cancer detection. Most public skin cancer image datasets are comprised of dermoscopic photos and are limited by selection bias, lack of standardization, and lend themselves to development of algorithms that can only be used by skilled clinicians. The SLICE-3D ("Skin Lesion Image Crops Extracted from 3D TBP") dataset described here addresses those concerns and contains images of over 400,000 distinct skin lesions from seven dermatologic centers from around the world. De-identified images were systematically extracted from sensitive 3D Total Body Photographs and are comparable in optical resolution to smartphone images. Algorithms trained on lower quality images could improve clinical workflows and detect skin cancers earlier if deployed in primary care or non-clinical settings, where photos are captured by non-expert physicians or patients. Such a tool could prompt individuals to visit a specialized dermatologist. This dataset circumvents many inherent limitations of prior datasets and may be used to build upon previous applications of skin imaging for cancer detection.

POT1 and multiple primary melanomas: the dermatological phenotype.

POT1 is the second most frequently reported gene (after CDKN2A) in familial melanoma. Pathogenic variants are associated with earlier onset and/or multiple primary melanomas (MPMs). To date, POT1 phenotypical reports have been largely restricted to associated malignancies, and description of the dermatological landscape has been limited. We identified 10 variants in n=18 of 384 (4.7%) unrelated individuals (n=13 MPMs; n=5 single primary melanomas) of European ancestry. Five variants were rare (minor allele frequency <0.001) or novel (two loss-of-function (LOF), one splice acceptor and two missense) and were predicted to be functionally significant, in five unrelated probands with MPMs (≥3 melanomas). We performed three-dimensional total body photography on both individuals with confirmed pathogenic LOF variants to characterise the dermatological phenotype. Total body naevus counts (≥2 mm diameter) were significantly higher (p=7.72×10-12) in carriers compared with a control population. Majority of naevi were on the probands' back and lower limb regions, where only mild to moderate ultraviolet (UV) damage was observed. Conversely, the head/neck region, where both probands exhibited severe UV damage, had comparably fewer naevi. We hypothesise that carriage of functionally significant POT1 variants is associated with increased naevus counts generally, and naevi >5 mm in diameter specifically and the location of these are independent of UV damage.

A protocol for annotation of total body photography for machine learning to analyze skin phenotype and lesion classification.

Introduction: Artificial Intelligence (AI) has proven effective in classifying skin cancers using dermoscopy images. In experimental settings, algorithms have outperformed expert dermatologists in classifying melanoma and keratinocyte cancers. However, clinical application is limited when algorithms are presented with 'untrained' or out-of-distribution lesion categories, often misclassifying benign lesions as malignant, or misclassifying malignant lesions as benign. Another limitation often raised is the lack of clinical context (e.g., medical history) used as input for the AI decision process. The increasing use of Total Body Photography (TBP) in clinical examinations presents new opportunities for AI to perform holistic analysis of the whole patient, rather than a single lesion. Currently there is a lack of existing literature or standards for image annotation of TBP, or on preserving patient privacy during the machine learning process. Methods: This protocol describes the methods for the acquisition of patient data, including TBP, medical history, and genetic risk factors, to create a comprehensive dataset for machine learning. 500 patients of various risk profiles will be recruited from two clinical sites (Australia and Spain), to undergo temporal total body imaging, complete surveys on sun behaviors and medical history, and provide a DNA sample. This patient-level metadata is applied to image datasets using DICOM labels. Anonymization and masking methods are applied to preserve patient privacy. A two-step annotation process is followed to label skin images for lesion detection and classification using deep learning models. Skin phenotype characteristics are extracted from images, including innate and facultative skin color, nevi distribution, and UV damage. Several algorithms will be developed relating to skin lesion detection, segmentation and classification, 3D mapping, change detection, and risk profiling. Simultaneously, explainable AI (XAI) methods will be incorporated to foster clinician and patient trust. Additionally, a publicly released dataset of anonymized annotated TBP images will be released for an international challenge to advance the development of new algorithms using this type of data. Conclusion: The anticipated results from this protocol are validated AI-based tools to provide holistic risk assessment for individual lesions, and risk stratification of patients to assist clinicians in monitoring for skin cancer.

MITF E318K: A rare homozygous case with multiple primary melanoma.

MITF E318K moderates melanoma risk. Only five MITF E318K homozygous cases have been reported to date, one in association with melanoma. This novel report uses 3D total-body-photography (TBP) to describe the dermatological phenotype of a homozygous MITF E318K individual. The case, a 32-year-old male, was diagnosed with his first of six primary melanomas at 26 years of age. Five melanomas were located on the back and one in the groin. Two were superficial spreading. Three arose from pre-existing naevi and one was a rare naevoid melanoma. 3D-TBP revealed a high naevus count (n = 162) with pigmentation varying from light to dark. Most naevi generally (n = 90), and large (>5 mm diameter) and clinically atypical naevi specifically were located on the back where sun damage was mild. In contrast, naevi count was low (n = 25 total) on the head/neck and lower limbs where sun damage was severe. Thus, melanoma location correlated with naevi density, rather than degree of sun damage. In addition to the MITF E318K homozygosity, there was heterozygosity for four other moderate-risk variants, which may contribute to melanoma risk. Further research is warranted to explore whether melanomas in E318K heterozygous and other homozygotes coincide with regions of high naevi density as opposed to sun damage. This could inform future melanoma screening/surveillance.

The MC1R r allele does not increase melanoma risk in MITF E318K carriers.

BACKGROUND: Population-wide screening for melanoma is not cost-effective, but genetic characterization could facilitate risk stratification and targeted screening. Common Melanocortin-1 receptor (MC1R) red hair colour (RHC) variants and Microphthalmia-associated transcription factor (MITF) E318K separately confer moderate melanoma susceptibility, but their interactive effects are relatively unexplored. OBJECTIVES: To evaluate whether MC1R genotypes differentially affect melanoma risk in MITF E318K+ vs. E318K- individuals. MATERIALS AND METHODS: Melanoma status (affected or unaffected) and genotype data (MC1R and MITF E318K) were collated from research cohorts (five Australian and two European). In addition, RHC genotypes from E318K+ individuals with and without melanoma were extracted from databases (The Cancer Genome Atlas and Medical Genome Research Bank, respectively). χ2 and logistic regression were used to evaluate RHC allele and genotype frequencies within E318K+/- cohorts depending on melanoma status. Replication analysis was conducted on 200 000 general-population exomes (UK Biobank). RESULTS: The cohort comprised 1165 MITF E318K- and 322 E318K+ individuals. In E318K- cases MC1R R and r alleles increased melanoma risk relative to wild type (wt), P < 0.001 for both. Similarly, each MC1R RHC genotype (R/R, R/r, R/wt, r/r and r/wt) increased melanoma risk relative to wt/wt (P < 0.001 for all). In E318K+ cases, R alleles increased melanoma risk relative to the wt allele [odds ratio (OR) 2.04 (95% confidence interval 1.67-2.49); P = 0.01], while the r allele risk was comparable with the wt allele [OR 0.78 (0.54-1.14) vs. 1.00, respectively]. E318K+ cases with the r/r genotype had a lower but not significant melanoma risk relative to wt/wt [OR 0.52 (0.20-1.38)]. Within the E318K+ cohort, R genotypes (R/R, R/r and R/wt) conferred a significantly higher risk compared with non-R genotypes (r/r, r/wt and wt/wt) (P < 0.001). UK Biobank data supported our findings that r did not increase melanoma risk in E318K+ individuals. CONCLUSIONS: RHC alleles/genotypes modify melanoma risk differently in MITF E318K- and E318K+ individuals. Specifically, although all RHC alleles increase risk relative to wt in E318K- individuals, only MC1R R increases melanoma risk in E318K+ individuals. Importantly, in the E318K+ cohort the MC1R r allele risk is comparable with wt. These findings could inform counselling and management for MITF E318K+ individuals.

Artificial Intelligence for the Classification of Pigmented Skin Lesions in Populations with Skin of Color: A Systematic Review.

BACKGROUND: While skin cancers are less prevalent in people with skin of color, they are more often diagnosed at later stages and have a poorer prognosis. The use of artificial intelligence (AI) models can potentially improve early detection of skin cancers; however, the lack of skin color diversity in training datasets may only widen the pre-existing racial discrepancies in dermatology. OBJECTIVE: The aim of this study was to systematically review the technique, quality, accuracy, and implications of studies using AI models trained or tested in populations with skin of color for classification of pigmented skin lesions. METHODS: PubMed was used to identify any studies describing AI models for classification of pigmented skin lesions. Only studies that used training datasets with at least 10% of images from people with skin of color were eligible. Outcomes on study population, design of AI model, accuracy, and quality of the studies were reviewed. RESULTS: Twenty-two eligible articles were identified. The majority of studies were trained on datasets obtained from Chinese (7/22), Korean (5/22), and Japanese populations (3/22). Seven studies used diverse datasets containing Fitzpatrick skin type I-III in combination with at least 10% from black Americans, Native Americans, Pacific Islanders, or Fitzpatrick IV-VI. AI models producing binary outcomes (e.g., benign vs. malignant) reported an accuracy ranging from 70% to 99.7%. Accuracy of AI models reporting multiclass outcomes (e.g., specific lesion diagnosis) was lower, ranging from 43% to 93%. Reader studies, where dermatologists' classification is compared with AI model outcomes, reported similar accuracy in one study, higher AI accuracy in three studies, and higher clinician accuracy in two studies. A quality review revealed that dataset description and variety, benchmarking, public evaluation, and healthcare application were frequently not addressed. CONCLUSIONS: While this review provides promising evidence of accurate AI models in populations with skin of color, the majority of the studies reviewed were obtained from East Asian populations and therefore provide insufficient evidence to comment on the overall accuracy of AI models for darker skin types. Large discrepancies remain in the number of AI models developed in populations with skin of color (particularly Fitzpatrick type IV-VI) compared with those of largely European ancestry. A lack of publicly available datasets from diverse populations is likely a contributing factor, as is the inadequate reporting of patient-level metadata relating to skin color in training datasets.

Improving Artificial Intelligence-Based Diagnosis on Pediatric Skin Lesions.

Artificial intelligence algorithms to classify melanoma are dependent on their training data, which limits generalizability. The objective of this study was to compare the performance of an artificial intelligence model trained on a standard adult-predominant dermoscopic dataset before and after the addition of additional pediatric training images. The performances were compared using held-out adult and pediatric test sets of images. We trained two models: one (model A) on an adult-predominant dataset (37,662 images from the International Skin Imaging Collaboration) and the other (model A+P) on an additional 1,536 pediatric images. We compared performance between the two models on adult and pediatric held-out test images separately using the area under the receiver operating characteristic curve. We then used Gradient-weighted Class Activation Maps and background skin masking to understand the contributions of the lesion versus background skin to algorithm decision making. Adding images from a pediatric population with different epidemiological and visual patterns to current reference standard datasets improved algorithm performance on pediatric images without diminishing performance on adult images. This suggests a way that dermatologic artificial intelligence models can be made more generalizable. The presence of background skin was important to the pediatric-specific improvement seen between models. Our study highlights the importance of carefully curated and labeled data from diverse inputs to improve the generalizability of AI models for dermatology, in this case applied to dermoscopic images of adult and pediatric lesions to improve melanoma detection.

Body Site Distribution of Acquired Melanocytic Naevi and Associated Characteristics in the General Population of Caucasian Adults: A Scoping Review.

The number of melanocytic naevi is a major risk factor for melanoma. The divergent pathway hypothesis proposes that the propensity for naevus proliferation and malignant transformation may differ by body site and exposure to ultraviolet (UV) radiation. This scoping review aimed to summarise the evidence on the number and distribution of naevi (≥ 2 mm) on the body overall and by individual anatomical sites in Caucasian adults, and to assess whether studies used the International Agency for Research on Cancer (IARC) protocol to guide naevus counting processes. Systematic searches of Embase and PubMed identified 661 potentially relevant studies, and 12 remained eligible after full-text review. Studies varied widely in their counting protocols, reporting of naevus counts overall and by body sites, and used counting personnel with differing qualifications. Only one study used the IARC protocol. Studies reported that the highest number of naevi was on the trunk in males and on the arms in females. Body sites which receive intermittent exposure to UV radiation had higher density of naevi. Larger naevi (≥ 5 mm) were detected mostly on body sites intermittently exposed to UV radiation, and smaller naevi (< 5 mm) on chronically exposed sites. Studies reported that environmental and behavioural aspects related to UV radiation exposure, as well as genetic factors, all impact body site and size distribution of naevi. This review found that to overcome limitations of the current evidence, future studies should use consistent naevus counting protocols. Skin surface imaging could improve the reliability of findings. An updated IARC protocol is required that integrates these emerging standards and technologies to guide reliable and reproducible naevus counting in the future.