Explainability agreement between dermatologists and five visual explanations techniques in deep neural networks for melanoma AI classification

Introduction

The use of deep convolutional neural networks for analyzing skin lesion images has shown promising results. The identification of skin cancer by faster and less expensive means can lead to an early diagnosis, saving lives and avoiding treatment costs. However, to implement this technology in a clinical context, it is important for specialists to understand why a certain model makes a prediction; it must be explainable. Explainability techniques can be used to highlight the patterns of interest for a prediction.

Methods

Our goal was to test five different techniques: Grad-CAM, Grad-CAM++, Score-CAM, Eigen-CAM, and LIME, to analyze the agreement rate between features highlighted by the visual explanation maps to 3 important clinical criteria for melanoma classification: asymmetry, border irregularity, and color heterogeneity (ABC rule) in 100 melanoma images. Two dermatologists scored the visual maps and the clinical images using a semi-quantitative scale, and the results were compared. They also ranked their preferable techniques.

Results

We found that the techniques had different agreement rates and acceptance. In the overall analysis, Grad-CAM showed the best total+partial agreement rate (93.6%), followed by LIME (89.8%), Grad-CAM++ (88.0%), Eigen-CAM (86.4%), and Score-CAM (84.6%). Dermatologists ranked their favorite options: Grad-CAM and Grad-CAM++, followed by Score-CAM, LIME, and Eigen-CAM.

Discussion

Saliency maps are one of the few methods that can be used for visual explanations. The evaluation of explainability with humans is ideal to assess the understanding and applicability of these methods. Our results demonstrated that there is a significant agreement between clinical features used by dermatologists to diagnose melanomas and visual explanation techniques, especially Grad-Cam.

COVLIAS 2.0-cXAI: Cloud-Based Explainable Deep Learning System for COVID-19 Lesion Localization in Computed Tomography Scans

Background: The previous COVID-19 lung diagnosis system lacks both scientific validation and the role of explainable artificial intelligence (AI) for understanding lesion localization. This study presents a cloud-based explainable AI, the “COVLIAS 2.0-cXAI” system using four kinds of class activation maps (CAM) models. Methodology: Our cohort consisted of ~6000 CT slices from two sources (Croatia, 80 COVID-19 patients and Italy, 15 control patients). COVLIAS 2.0-cXAI design consisted of three stages: (i) automated lung segmentation using hybrid deep learning ResNet-UNet model by automatic adjustment of Hounsfield units, hyperparameter optimization, and parallel and distributed training, (ii) classification using three kinds of DenseNet (DN) models (DN-121, DN-169, DN-201), and (iii) validation using four kinds of CAM visualization techniques: gradient-weighted class activation mapping (Grad-CAM), Grad-CAM++, score-weighted CAM (Score-CAM), and FasterScore-CAM. The COVLIAS 2.0-cXAI was validated by three trained senior radiologists for its stability and reliability. The Friedman test was also performed on the scores of the three radiologists. Results: The ResNet-UNet segmentation model resulted in dice similarity of 0.96, Jaccard index of 0.93, a correlation coefficient of 0.99, with a figure-of-merit of 95.99%, while the classifier accuracies for the three DN nets (DN-121, DN-169, and DN-201) were 98%, 98%, and 99% with a loss of ~0.003, ~0.0025, and ~0.002 using 50 epochs, respectively. The mean AUC for all three DN models was 0.99 (p < 0.0001). The COVLIAS 2.0-cXAI showed 80% scans for mean alignment index (MAI) between heatmaps and gold standard, a score of four out of five, establishing the system for clinical settings. Conclusions: The COVLIAS 2.0-cXAI successfully showed a cloud-based explainable AI system for lesion localization in lung CT scans.

Grad-CAM helps interpret the deep learning models trained to classify multiple sclerosis types using clinical brain magnetic resonance imaging

BACKGROUND

Deep learning using convolutional neural networks (CNNs) has shown great promise in advancing neuroscience research. However, the ability to interpret the CNNs lags far behind, confounding their clinical translation.

NEW METHOD

We interrogated 3 heatmap-generating techniques that have increasing generalizability for CNN interpretation: class activation mapping (CAM), gradient (Grad)-CAM, and Grad-CAM++. To investigate the impact of CNNs on heatmap generation, we also examined 6 different models trained to classify brain magnetic resonance imaging into 3 types: relapsing-remitting multiple sclerosis (RRMS), secondary progressive MS (SPMS), and control. Further, we designed novel methods to visualize and quantify the heatmaps to improve interpretability.

RESULTS

Grad-CAM showed the best heatmap localizing ability, and CNNs with a global average pooling layer and pretrained weights had the best classification performance. Based on the best-performing CNN model, called VGG19, the 95th percentile values of Grad-CAM in SPMS were significantly higher than RRMS, indicating greater heterogeneity. Further, voxel-wise analysis of the thresholded Grad-CAM confirmed the difference identified visually between RRMS and SPMS in discriminative brain regions: occipital versus frontal and occipital, or temporal/parietal.

COMPARISON WITH EXISTING METHODS

No study has examined the CAM methods together using clinical images. There is also lack of study on the impact of CNN architecture on heatmap outcomes, and of technologies to quantify heatmap patterns in clinical settings.

CONCLUSIONS

Grad-CAM outperforms CAM and Grad-CAM++. Integrating Grad-CAM, novel heatmap quantification approaches, and robust CNN models may be an effective strategy in identifying the most crucial brain areas underlying disease development in MS.