Melanoma is one of the top-5 most common cancers in Switzerland with a standardized incidence ratio per 100,000 inhabitants of 29.8 for men and 24.7 for women [1]. Longitudinal data acquired since 1989 show a linear 100% increase of the standardized incidence ratio for men [1] in the last 30 years. Unfortunately, this is a worldwide trend. According to Arnold et al [2], melanoma incidence and deaths are expected to increase 50% and 68%, respectively, by 2024. However, it is known that if diagnosed on time, skin cancer can be cured with a simple surgical procedure, dramatically increasing the survival rate. The rapid increase in cases every year has not been followed by an increase in the available number of dermatologists. This causes the system to operate inefficiently with increasing waiting times to get access to specialist consultation. To cope with such a situation, a study [3] proposed to provide specific training in skin cancer diagnosis to general practitioners to improve their competence in such cases. The outcome of that study showed a positive impact during a limited period after the initial training, but eventually, the accuracy dropped again after one year. Thus, it is necessary to explore long-term solutions that can support dermatologists to face this pandemic in a reliable way.

The use of deep learning has raised substantial interest in dermatology. Seminal studies such as Esteva et al [4] showed how deep learning algorithms can outperform board-certified dermatologists in certain dermoscopic image triage tasks. More recent studies broadened the scope by developing algorithms for automatic screening and ugly-duckling characterization in wide surface images [5]. Despite such encouraging applications and results, the transition from academic research studies to real-world application is only slowly being addressed. Different surveys showed a favorable position from dermatologists [6] as well as patients [7] with respect to the introduction of artificial intelligence (AI) in routine consultations. However, interpretability [8] and the need of a specialist to supervise the outcome are issues that concern clinical users and patients. When exploring the option of implementing such support tools in real-world consultation, there are 2 main aspects to monitor: on the one hand, diagnosis accuracy and, in the other, diagnosis confidence, since both play a key role. For example, the overdiagnosis of benign lesions as malignant lesions and diagnoses with low confidence cause unnecessary surgeries. A frequent scenario in clinical practice is the surgical removal of atypical-looking benign lesions. Even in the hands of experts, there are 5 benign lesions removed for every 1 melanoma [9], and in the hands of nonexperts, this increases exponentially. Another important fact that needs to be carefully considered is the bias that can be introduced in the user’s decision-making. However, one of the main limitations in the adoption of these tools for nonexpert users comes from the interpretability of such tools, which does not transmit confidence in the predictions even if the demonstrated accuracy is high.

Content-based image retrieval (CBIR) is a powerful tool in medical practice that proposes similar cases to the ones under study, thus mimicking an automatized bibliography search. Dermatology is not an exception, since large data sets of images are traditionally available. Early applications of CBIR in skin lesion categorization relied on the comparison retrieved from text and annotations [10] and progressed to consider colors and shapes [11,12]. However, with the introduction of convolutional neural networks (CNNs), CBIR is now using them as the backbone for feature extraction [13-17]. A pilot study [16] concluded that CBIR was perceived as easy to use and engaging; however, trust still has to be gained for routine use. Furthermore, in another high-impact study [17] where the impact of using CBIR as a support tool was evaluated, one of the conclusions was that users tend to prefer other AI approaches such per-class probability in the long term. The authors suggested exploring other CBIR architectures that might overcome this issue. One possible reason is the lack of interpretability of the retrieval process. To overcome this limitation, recent studies [18-20] evaluated the option of introducing interpretability methods such as saliency maps to guide the CNN feature extraction. These newly proposed architectures were tested using radiography data sets, which present substantial differences with dermoscopic imaging. For this reason, in this study, we proposed an updated version of the algorithm by Silva et al [18] for the specific needs of skin lesion diagnosis, which we refer to as saliency-enhanced CBIR (SE-CBIR).

The benefits of SE-CBIR in routine dermatologic consultations are as follows:

For the training, validation, and testing of our algorithm, the HAM10000 data set [21,22] was used. It consists of a total of 10,015 labeled dermoscopic images belonging to 7 different categories of pigmented lesions. The ground truth was determined in more than 50% of images by histopathology, and the other half were either confirmed by follow-up examinations, expert consensus, or in vivo confocal microscopy [22]. Follow-up images of lesions were removed during the retrieval process to maximize the diversity. The HAM10000 data set designations for the 7 different classes were kept. The description for each category is as follows: actinic keratosis, squamous cell carcinoma, and Bowen disease (akiec); basal cell carcinoma (bcc); benign keratosis (bkl; solar lentigo, seborrheic keratosis, and lichen planus-like keratosis); dermatofibroma (df); melanoma (mel); melanocytic nevi (nv); and vascular lesions (vasc; angiomas, angiokeratomas, pyogenic granulomas, and hemorrhage). The data set is highly unbalanced, with 67% of the total number of lesions corresponding to melanocytic nevi and 11% to melanoma. To counteract such imbalance, a multiclass focal loss function was used during training.

In this study, we used a public data set, namely HAM10000 [21,22]. This data set is routinely used by research studies and is available under a Creative Commons Attribution-NonCommercial 4.0 International Public License. Therefore, the use of this data set is not subject to ethics board approval.

CBIR algorithms are designed to retrieve images from a data set that are related to the image under study. Their performance is boosted by introducing CNNs for feature extraction and representation. We chose EfficientNetB4 [23] as the backbone for our classifier. The choice was driven by the fact that ensemble models combining different versions of EfficientNet were used by the top teams in International Skin Imaging Collaboration (ISIC) competitions [24]. We profited from transfer learning by initializing the weights using “ImageNet” default values. The EfficientNetB4 top layer was removed and replaced by average pooling, batch normalization, and 2 combinations of a dropout and a dense layer. The output layer features a softmax activation function. We extracted the deep features just after the last convolutional layer and before the classification layers.

The traditional CBIR approach, as depicted in Figure 1, includes 3 steps: step 1, a CNN classifier using the HAM10000 data set is trained; step 4, extracted features from the query image are compared to those of each image in the retrieval set; and step 5, images are ranked according to similarity.

In clinical imaging, the relevant information is usually spatially constrained, so new approaches such as that in Silva et al [18] propose to add a fine-tuning stage, guiding the training using saliency maps to enhance the retrieval performance. With the emerging field of explainable AI, a variety of saliency methods have been developed, such as vanilla gradient [25], SmoothGrad [26], and integrated gradients [27]. Saliency methods aim to create a map highlighting the pixels that are relevant for the network’s classification of a particular input image. We chose vanilla gradient for various reasons: it is model agnostic, simple to integrate, interpretable, and consistent. In Silva et al [18], the superiority of latent representations derived from saliency maps over those directly generated from the input x-ray images was demonstrated, where latent representations of saliency maps were used for medical image retrieval purposes and in sample selection for active learning.

Due to the substantial differences between the x-ray images and our dermoscopic data set, we designed a second fine-tune stage that input the original image (3 RGB channels) plus the saliency map computed in the first step. This was done to avoid excessive loss of information included in the original colored dermoscopic images in contrast with the approach in Silva et al [18]. As shown in Figure 1, the traditional CBIR approach is comprised of 3 steps: the training of a skin lesion classifier (step 1), followed by deep feature extraction for every single image (step 4), and finally image retrieval according to similarity score (step 5), whereas SE-CBIR comprises 5 steps, adding the computation of saliency maps (step 2) and the training of a 4-channel classifier (step 3) [28]. It is worth it noting that in step 3, the input and first convolutional layer are expanded to a 4-channel input to input the saliency maps together with the original image.

Standard data augmentation techniques from the albumentations libraries [29] were applied during training to avoid overfitting. These transformations include geometric augmentations and noise, distortion, brightness and contrast, and color modifications. The data augmentation is applied on the fly during each epoch. All images were resized and randomly cropped to the expected EfficientNetB4 input resolution (380 × 380). Additionally, coarse dropout to enhance regularization was applied. During the training of SE-CBIR (fine-tuning stage), no data augmentation, other than random flips and rotations, was performed.

For training, validation, and test purposes, the data set was randomly stratified in an 80:10:10 split, ensuring the same class distribution in all subsets. During training, an adaptative learning rate (α) approach was chosen to allow different α values for the pretrained EfficientnetB4 layers and the added layers, aiming to adapt the parameters of the pretrained layers just slightly or not at all compared to the new layers. A learning rate schedule was set up, with an initial ramp-up during the first epochs that aims to keep the learned features. After the ramp-up, the learning rate decays exponentially. Exponential decay is a widely used learning rate scheduling method to improve convergence. The adaptability of this approach reduces the time required to train neural networks and makes a neural model scalable, as they can adapt to structure and input data at any point in time while being trained.

For step 1, the learning rate was set to a low value (α=5 × 10^–5) to profit from prelearned features, whereas for the additional layers, the α was set to .01. This approach focuses the training on the classification layers. In this second step, we added an additional channel to the original input size, creating 4D images including the 3 RGB channels plus the 1D saliency maps. An adaptive learning rate approach similar to traditional CBIR was applied. However, the adapted layers were found to perform better at being trained with a 100-times larger (α=1 × 10^–3) learning rate than that in the first step. The learning rate was kept low at α=1 × 10^–5 for the pretrained layers to profit again from transfer learning, and the same number of epochs and loss function were used.

The multicategorical focal loss function [30] was chosen for training. This function adds a term to the cross-entropy loss, improving performance in imbalanced data sets such as HAM10000. This is achieved through down weighting, which reduces the influence of easy examples on the loss function, resulting in more attention to hard ones. With a softmax activation function in the last layer, the focal loss ℒ_focal for each sample can be derived by:

The parameters α_i and γ define the weights on this additional term, whereas y*_i represents the softmax prediction value of an input.

In CBIR and image classification-based models, high-level image visuals are represented in the form of feature vectors that consists of numerical values. Image retrieval is based on the comparison of features extracted from the images. These vector features are then compared among the different image features to search and rank the “closest” ones. We chose cosine similarity to compare the features’ latent representation due to its adequacy of handling high-dimensional vectors (1792 × 1 in our case). The mathematical representation is as follows:

where A and B are the feature vectors of the query image and each of the images in the labeled data set. The lower the value of S_c, the closer both images are in terms of features extracted.

To evaluate the retrieval performance and compare the traditional CBIR algorithm against SE-CBIR, we used the retrieval precision for k retrieved images (P@k) or cut-off value metric:

Similarly, for multiclass retrieval, we defined the average precision for k retrieved images (AP@k):

with M being the total number of classes.

To evaluate the impact of this type of decision support, a reader study was designed within the University Hospital Zurich Dermatology Clinic. A total of 9 participants were recruited based on their willingness and availability to participate: 1 expert on pigmented skin lesions with more than 25 years of experience and 8 residents from the clinic with 1 to 5 years of experience. However, all of them had reasonable experience with melanoma diagnosis using dermoscopic imaging.

To evaluate the impact of the diagnosis support tool, the reader study was divided into 2 tasks. In task 1, volunteers were asked to provide a diagnosis on 100 randomly chosen dermoscopic images extracted from our HAM10000 test set. In parallel, they were also asked to provide their level of confidence in their diagnosis on a 5-point Likert scale [31]. After a break period of 1 day, the participants underwent task 2. They were asked to rediagnose the same set of images while being supported with the 6 “closest” images proposed by the retrieval algorithm that are characterized by their associated ground-truth label, as shown in Figure 2. Note that the saliency maps were not presented to the participants along with the retrievals; however, in a real-world implementation, they would be available. To minimize bias, the single images were rotated by 180° with respect to the original orientation. Additionally, each participant was presented with a different, randomly selected set of images for the evaluation. As in task 1, the volunteers were asked to provide a diagnosis and their confidence level. To facilitate participation, a web survey was created ad hoc for the evaluation process, allowing the users to examine the lesions on high-quality screens and providing flexibility on when and where to perform the evaluation.

An example of a saliency map calculated using the vanilla gradients is shown in Figure 3 (center). Additionally, an overlay of the original image and the saliency map is shown (right) for illustration purposes. To minimize the loss of information during the fine-tuning training step, the original image (3 RBG channels) and the corresponding saliency map (1 channel) are fed as a 4-channel input to the CNN.

Retrieval precision comparing the traditional CBIR algorithm versus SE-CBIR is summarized in Table 1. The saliency-enhanced approach improves both the per-class prediction (P@k) and the average (AP@k) for all values k of retrieved images. The improvement in AP@k increases with k, from 7% (k=1) to a maximum of 18% (k=9). For a single class, the largest difference was found for class df where the prediction accuracy was almost doubled. The only class that did not experience a significant improvement was akiec. Both classes df and akiec were underrepresented with only 1% and 2% of the total retrieval data set. In the case of class df, SE-CBIR identified additional features leading to a significant improvement, whereas for class akiec, it could be that the limited samples available for class akiec presented similar appearances as solar lentigo or seborrheic keratosis and larger statistics are needed. For all other classes, P@k was >0.8 for k=1.

The outcome of the clinical evaluation by the 9 volunteers is summarized in Table 2. Diagnosis accuracy for the SE-CBIR algorithm was computed by majority voting among the k=6 retrieved images, which reached 89% on average for all evaluations. In task 1, participants recorded the lowest accuracy at 38%, with an average of 62.2%. The performance was not uniform with σ=12.7 points. As expected, the board-certified dermatologists performed at the same level as the algorithm. In task 2, a significant improvement in diagnosis accuracy was observed for all participants. The average accuracy increased by 22.7 points to 84.9%, which also helped to reduce diagnosis spread, bringing it down to σ=7.1 points. Since each participant was presented with a different image data set, to evaluate the consistency of their selection, the Cohen κ coefficient for each participant for tasks 1 and 2 was calculated, as well as the average among them. The results showed that agreement between raters improved with the support of AI, going from fair agreement (k_average=0.35) in task 1 to substantial agreement (k_average=0.66) in task 2. In addition, the 2 best-performing participants in task 1 were able to outperform the SE-CBIR majority-voting prediction when provided with retrieved images. It is worth noting that the HAM10000 data set is composed of hand-picked lesions whose diagnosis can be challenging without access to the patient’s context. Regarding the average diagnosis confidence, an improvement from 3.11 to 3.86 was found (+24%). With the support of the retrieved cases, the confidence in the correct diagnosis increased from 3.35 to 4.03. Indeed, most (6/9, 67%) of the participants expressed a confidence level above 4, both in absolute confidence and confidence for correct diagnoses, whereas in task 1, only 1 participant showed that level of confidence. On the other hand, the confidence in incorrectly predicted lesions increased much less (2.72 to 2.90).

Table 2 shows the performance of the retrieval algorithm with an average accuracy of 89.2%. This result is substantially higher than the largest reader study using the ISIC18 data set published by Tschandl et al [17]; however, it used a different test set for the evaluation.

A closer look at the predictions per class is shown in Figure 4 with the help of confusion matrices. Each row represents the total number of instances of a given class, whereas each column represents the total number of instances predicted for each class. Figure 4 (right) shows a substantial improvement of per-class accuracy with the help of SE-CBIR–retrieved images.

Class nv represented almost 80% of the test data set, and the SE-CBIR results show that almost 25% of the diagnoses were reconsidered into the correct class. Similarly, for class bkl, incorrect diagnoses were corrected in task 2 improving the accuracy by 17%. For class mel, the accuracy improved to 79% for correct melanoma cases. In total, just 20 correct decisions of task 1 were changed to an incorrect diagnosis in task 2, whereas 224 misclassified lesions from task 1 were correctly classified in task 2. This is an important indication that the algorithm does not lead to overconfident misclassifications but rather improves the number of correct diagnoses and their confidence—although we should note that 12 of the incorrectly overturned diagnoses agreed with the algorithm’s majority voting. Those cases probably would require additional patient-context information for a better evaluation. From a skin management point of view, melanoma overdiagnosis is a conservative approach where a negative biopsy might be justified. Regarding melanoma (mel) diagnosis, in task 1, a total of 21 cases were underdiagnosed (33% of the 64 total mel cases), whereas in task 2, this value decreased to 14 (22%) out of 64 cases. Regarding overdiagnosis in task 1, a total of 171 nonmelanoma cases (where 144 were nv cases) were misclassified into mel, whereas in task 2, this value decreased to 81 cases (with 65 nv cases).