Paige Artificial Intelligence-Assisted Nevus Classification: Comparing Measurement Methods to Support Pathologic Diagnosis

Objective: Observe if an artificial intelligence (AI) algorithm can serve as an assistive tool for pathologists in differentiating between benign, varying degrees of atypical, and malignant skin nevi in cases that are microscopically similar or ambiguous.

Methods: A cross-sectional study was conducted on nevi cases received from JIT Laboratories, a single-origin laboratory in Irvine, California. These cases were classified as benign (n=11), mildly atypical nevus (n=6), moderately atypical nevus (n=5), severely atypical nevus (n=6), melanoma in situ (MIS, n=7), and malignant melanoma (MM, n=5). Slides consisting of multiple sections of nevi were digitized via PANNORAMIC 250 and PANNORAMIC1000 by Epredia, anonymized through SlideMaster, and uploaded to Paige, where the PAIGE PanCancer Detect system then created highlighted overlays that the algorithm determined malignant potential. These overlays were measured using the Paige platform ruler function. For more typically shaped overlays, we measured the longest longitudinal length from left-to-right, end-to-end; for more irregular/difficult to measure overlays, we measured the longitudinal length by measuring the length resulting from connecting the tips of the rete ridges. Four methods of measurement and calculations were conducted. A section was considered a group of tissue on the slide. A segment was considered a slice of tissue from one tissue group on the slide. Method 1: the longest sum of highlighted overlays in a single segment within a single section on a slide; Method 2: the results from Method 1 divided by the gross biopsy specimen size; Method 3: the longest sum of all highlighted overlays in a single section on a slide; and Method 4: the longest individual highlighted region. Kruskal Wallis test was used to compare the six nevi groups, followed by Dunn's post-hoc test for pairwise comparisons. P < 0.05 was considered statistically significant for Kruskal Wallis, while a corrected α (Bonferroni correction method) of 0.003333 was used to determine significance for Dunn’s.

Results: A total of 40 cases were assessed via PAIGE PanCancer Detect software, which highlighted 612 overlays. ROI measurements differed significantly among the six nevi categories (Kruskal-Wallis, p = 0.002523). The median ROI length for benign was 0.99 mm (IQR 0.35-1.735), mildly atypical nevus was 1.36 mm (IQR 1.01-1.79), moderately atypical nevus was 1.77 mm (IQR 1.28-2.16), severely atypical nevus was 1.77 mm (IQR 0.9-3.33), MIS was 4.06 mm (IQR 3.03-7.04), and MM was 4.94 mm (IQR 3.49-5.12).

Conclusion: Our preliminary results suggest that there is a statistically significant difference between the ROI of benign nevi vs. MM (p = 0.0014), and benign nevi vs. MIS (p = 0.00069). When accounting for the size of the gross biopsied lesion, there is a statistically significant difference between benign nevi and MM (p = 0.0019). This suggests that an artificial intelligence software can be used to differentiate between benign and malignant melanoma. However, further work and a greater sample size are needed to evaluate whether there are differences between microscopically similar cases, such as mildly atypical nevi vs moderately atypical nevi or moderately atypical nevi vs severely atypical nevi.