Hybrid Deep Learning and Handcrafted Feature Analysis for Melanoma Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current dermoscopy imaging methods, including those used by dermatologists, often miss cases of melanoma and other skin cancers, with existing computer techniques like deep learning and machine vision showing higher diagnostic accuracy but still not achieving optimal results in classification and segmentation of dermoscopy images.
Innovation Solution
A system and method that combines handcrafted feature detection with deep learning techniques, utilizing a segmenter, handcrafted feature component, and deep learning feature component to preprocess and analyze dermoscopy images, including median color splitting, vessel detection, atypical pigment network detection, and salient point detection, to enhance image segmentation and classification accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If dermatologists use dermoscopy with their training and experience, then diagnostic capability is improved, but diagnostic accuracy is lower than computer techniques
Solution Approach 1:
The patent combines handcrafted feature detection (segmentation, color analysis, vessel detection, pigment network detection) with deep learning classification to create a hybrid system that achieves superior diagnostic accuracy (AUC > 0.90) compared to either approach alone, resolving the contradiction between human expert capability and computer technique performance
2Measurement precision
If standalone deep learning techniques are used for dermoscopy image analysis, then automation is improved, but diagnostic accuracy is lower than the combined approach
Solution Approach 1:
The patent segments the image analysis process into distinct handcrafted components (lesion segmentation, color region analysis, vessel detection, pigment network detection) that preprocess and prepare features for the deep learning classifier, improving overall accuracy while maintaining automation
3Measurement precision
If simple segmentation is applied to dermoscopy images, then processing speed is improved, but segmentation accuracy is insufficient for optimal classification
Solution Approach 1:
The patent applies preliminary handcrafted image processing operations (thresholding for segmentation, color splitting, vessel detection, pigment network detection) before deep learning classification to extract and prepare critical features, improving classification accuracy while maintaining efficient processing through optimized algorithms
Data Source
AI summary
A system for identifying melanoma and other skin cancer in a dermoscopy image comprises: an image analyzer having at least one processor that instantiates at least one component stored in a memory, the at least one component comprising: a segmenter configured to segment a lesion from the rest of the image, a handcrafted feature component including: a median color splitting model for separating the image into a plurality of color regions, a vessel detection model for detecting elevated vascularity, an atypical pigment network detection model for identifying a pigment network whose structure varies in size and shape, a salient point detection model for detecting salient points based on an intensity plane of the image, a color detection model for detecting at least one of a white area, a pink shade, a pink blush, and a semi-translucency, a hair detection model for characterizing detected hairs and ruler marks, an outside model that finds the above model features on non-dark-corner areas outside the segmented area, and a classifier configured to provide a first analysis result.