Abstract
Among Computer-Aided Diagnosis (CAD) systems applied to melanoma detection on dermoscopic images, neural network approach has shown high performance. However, the experimental conditions of these studies are far from the clinical practice, errors can occur and are sometimes obvious for a dermatologist. However, the lack of explainability of the learning and classification process make nterpretation of these errors challenging. In contrast, classical CAD approaches required several steps for their development ; and the developer can extract features guided by dermatological concepts which will implement the model. However, most have been developed using a combination of criteria, notably ABCDE (A: asymmetry, B: irregular edges, C: inhomogeneous color, D: diameter greater than 6 mm and E: evolution), whereas dermatologists, when clinically examining a pigmented lesion, do not use a combination of criteria, but unconsciously use a more global cognitive analysis.
Indeed, they probably distinguish melanomas by concepts they perceive as disorder, chaos or lack of overall symmetry of the lesion. To our knowledge, these concepts have not been integrated yet into CAD systems. Our aim was to develop models based on classical approaches incorporating concepts inspired by a global cognitive analysis, to compare their performance with neural approach, and finally to see whether combining the two approaches would improve the performance of neural approaches and then avoid obvious errors. To achieve that, we used the International Skin Imaging Collaboration’s (ISIC) public dataset (BDD) of dermoscopic images from 2019, after improving and standardizing it by limiting bias using image processing technique. We thus obtained a dataset containing 1,533 melanomas and 6,124 nevi.
We hypothesized that the clinical diagnosis of melanoma by a dermatologist is based on the human expert’s perception of concepts such as “disorder” regarding the color distribution or shades of color traducing by the “chaos” i.e. significant color contrasts of juxtaposed regions of the lesion, and finally “asymmetry”. We attempted to model all these concepts mathematically.
We first developed a supervised model with the K-Nearest-Neighbors (KNN) classifier, characterizing the lesion on an order/disorder scale thus the concept of melanoma disorder thanks to the entropy descriptor on several color spaces. The performance of the “disorder model” proved equivalent results to a neural approach developed based on the ResNet-50 architecture. We then developed an algorithm characterizing the lesion on a “chaos”
scale using graph modeling of segmented lesions integrated into six classifiers. Then, we developed a convolution neural network (CNN) based on the EfficientNet architecture on the same dataset. We then combined our two approaches based on the final predictions of each model by calculating the mean of their predictions. We developed an algorithm based on the degree of asymmetry integrating two forms of symmetry: central symmetry and axial symmetry. Features from four color spaces were used to train an Artificial Neural Network. We then developed an ensemble method based on multiple CNNs on the same dataset, using the EfficientNet architecture. We then combined our two approaches based on the final predictions of each model calculating the mean of their predictions. Finally, we fuse all our models using another method, taking into account our different algorithms inspired by the clinician’s global cognitive approach, and compared them with the CNNs’ performance. The results obtained show that our approach provided equivalent performance of a CNN and corrects the majority of the initial obvious errors made by the CNN.
This work highlights the algorithms’ diagnostic performance based on concepts such as “disorder”, “chaos” and “asymmetry”, and the potential contribution of concepts from dermatological practice and inspired by the global cognitive approach of the Dermatologist to improve the performance of CADs based on CNN methods and to reduce the so-called obvious errors for the clinician.
Committee:
- Caroline ROBERT, IGR, INSERM, University of Paris-Sud, Paris-Saclay (Chair of the Jury)
- Tu-Anh DUONG CHU, Ambroise Paré Hospital, AP-HP, University of Paris Saclay, Centrale Supélec, Industrial Engineering Laboratory (Thesis Examiner)
- Abd-Krim SEGHOUANE, School of Mathematics and Statistics, University of Melbourne (Thesis Examiner)
- Véronique DEL MARMOL, Erasme Hospital, Université Libre de Bruxelles (Examiner)
- Caroline GAUDY-MARQUESTE, CHU La Timone, CRCM, INSERM, Aix-Marseille University (Examiner)
- Rabah IGUERNAISSI, LIS, Aix-Marseille University (Examiner)
- Jean-Jacques GROB, CHU La Timone, CRCM, INSERM, Aix-Marseille University (PhD Supervisor)
- Djamal MERAD, LIS, Aix-Marseille University (PhD Co-Supervisor)