Computer-Based Detection of Colorectal Serrated Lesions: Digital Flatness, a Novel Metric Designed for Whole-Slide Images

Colorectal sessile serrated lesions (SSLs) and hyperplastic polyps (HPs) are characterized by sawtooth or stellate epithelial architecture. Distinguishing between SSLs and HPs is crucial as SSLs are precursors of colorectal carcinomas in 30% of cases, whereas HPs are likely precursors to SSLs. The differentiation of SSL from HP is primarily based on architectural features. Indeed, the hallmark of SSL is a substantial distortion of the typical crypt design and silhouette, which shows horizontal expansion along the muscularis mucosae and enlargement of the crypt base, especially in the lower third of the crypt. The ability to analyze digitized histologic images has led to innovative automated tissue analysis, thereby improving reproducibility and objectivity in pathologists' reports. Some recent studies explored colorectal cancer diagnosis and grading through automated quantitative analysis, but none of them focused on SSL detection. This study aimed to develop an automated method for SSL diagnosis by defining specific metrics to characterize their most common visual features. We developed a processing pipeline involving the automatic segmentation of all the tissue structures required for computing quantitative morphologic and architectural features, which allows detection of SSLs. In particular, we designed a novel metric, digital flatness, which numerically characterizes the parallelism of the gland's contour edges with the muscolaris mucosa profile. In a data set of 759 polyp glands, 41 of which were reported as SSLs by expert pathologists, our novel detection method achieved specificity of 92% and sensitivity of 83%, with accuracy of 92%. Our results represent a first approach to a simple, common, but still debated issue among gastrointestinal pathologists, thus providing valid support for the objective and standardized individuation of SSLs.