Enhancing skin cancer detection with explainable artificial intelligence: A customized extended deep U-shaped encoder decoder network approach

2.1 U-Net

The model proposed is a powerful architecture designed for the segmentation process using biomedical image samples, known as the U-net architecture (Sanjar et al., 2020). Our extended deep convolutional U-net architecture, depicted in Fig. 2 consists of three main blocks: encoding-decoding block (Wu et al., 2021) and attention gate mechanism function (Wu et al., 2021). The detailed configuration has been illustrated in Fig. 3(a). The block diagram of the proposed extended U-net in Fig. 3(b) displays a series of four encoding blocks defined by convolutional layers, which aim to capture the context of the input image feature maps. This block helps to reduce the spatial dimensions and feature maps while increasing depth. In the decoder block, the image progressively recovers the spatial resolution using transpose convolutions or up-sampling followed by convolution. The decoder block includes skipping connections from the encoder block to preserve high-resolution features through concatenation. Attention gates are added before merging features in the decoder. By focusing on relevant features and suppressing irrelevant regions, the attention gate block helps the model when dealing with input images of skin cancer. Gating signals are used in this model to emphasize important features passed through skip connections.

Close

Fig. 2.

Block diagram of proposed U-net architecture based on deep learning.

Export to PPT

Close

Fig. 3.

Provides a detailed architectural representation of the proposed model. (a) Illustrates the block diagram of individual components, including the encoding, decoding, and attention gate architectures, which play a crucial role in feature extraction and refinement. (b) Presents the overall structure of the proposed extended U-Net model, built upon a deep convolutional neural network, highlighting its enhanced capability for skin cancer segmentation and classification.

Export to PPT

2.2 Encoding (contracting path)

In proposed extended deep convolutional U-net architecture, the encoder block (Aboussaleh et al., 2023) is essential for capturing convolutional information and decreasing the spatial dimensions of the input image. It follows a neural network architecture like that of a CNN. In Fig. 3(a), a single encoding block includes a convolution layer, pooling layer, dropout layer, and feature doubling, resembling a CNN. The details and mathematical overview of these layers are as follows:

2.3 Decoding

A decoder (Kim et al., 2018) is defined as a functional block inside network layer configuration in Fig. 3(a), in the context of extended U-net architecture that is responsible for up-sampling. The up-sampling process is carried out by the encoder function block, which gathers and combines feature maps to create a segmentation map or an output with the same dimensions as the input sample, but possibly with a different number of channels. In our U-net architecture, the decoder block consists of four up-sampling blocks, each performing up-sampling followed by convolutional operations. We utilized transpose convolutions, also known as up-convolutions or deconvolutions, to expand the spatial dimensions in feature maps. Additionally, we incorporated the feature maps from the corresponding encoder block by concatenation, enabling the decoder to access both high-level and low-level features. Let denotes the i^th samples of input feature map to the decoder as X with dimension (${\text{U}}_{\text{in}}\times {\text{V}}_{\text{in}}\times {\text{C}}_{\text{in}}$) where, ${\text{U}}_{\text{in}}$ represent as height and ${\text{V}}_{\text{in}}$ as width of the feature maps. The channel number is denoted as ${\text{C}}_{\text{in}}$. After up-sampling, the output feature maps produce (${\text{U}}_{\text{in}}\times {\text{V}}_{\text{in}}\times {\text{C}}_{\text{in}}$) same dimension as feature map but the ${\text{C}}_{\text{in}}$ value might be changed. In our approach, up-sampling has been used as transposed convolution also known as fractionally stride convolution or deconvolution. In a mathematical way it is represented as follows:

where is the output feature map after upsampling and Conv2DTranspose represents the transpose convolution operation.

2.4 Attention gate

The model used four attention gate mechanism blocks (a1, a2, a3, and a4) connected with four decoding blocks (d1, d2, d3, and d4) in Fig. 3(b) individually in our proposed U-net architecture. Each attention gate functional block consists of a sequential series of three convolutional layers, up-sampling 2D, and batch normalization displayed in Fig. 3(a). Basically, the attention gate is used in neural networks, particularly in architectures like RNN (Recurrent Neural Network), CNN, and image segmentation U-net architecture. By following human attention, the attention gate mechanism is inspired. The attention gate consists of two main components as follows:

2.5 Grad-CAM

Gradient-weighted Class Activation Mapping, (Selvaraju et al., 2017) (Grad CAM) is a technique used for visualizing image feature maps. It makes use of the gradients of a target concept that flow from the final convolutional layer. This technique produces a feature map localization, which highlights the affected or targeted areas for original image segmentation with its predicted and processed image mask. Let’s consider V^th as the feature maps of the final convolutional layer, denoted as G^d. In gradient computation, r is denoted as the score for the class. Therefore, n^r with respect to the final convolutional layer feature maps G^d, defined as $\frac{\partial n^r}{\partial n^r}$. In the mathematical equation performed, the neuron importance weights $a_d^r$ with global pooling average on the gradients as follows:

where (i,j) index the spatial dimensions of feature maps and (M) denoted as pixel numbers in the feature map.

The Weighted combination of the feature maps:

Where to generate the heatmap $K_{Grad-CAM}^c$ is the resulting upscaled to input sample size of images. Usually, in linear combinations of maps, the ReLU function is used to keep features. On the level of interest, this function has a positive influence.

2.6 Proposed network architecture

In Fig. 3(b), an encoder block consists of two functional components denoted as ($p_n$, $c_n$), where $p_n$ represents the encoder output and $c_n$ indicates the input to the block. The attention-gating block is represented as $a_n$, while the decoding block is denoted as $d_n$. Each of these blocks, encoding, decoding, and attention gating comprises convolutional layers, max-pooling layers, upsampling layers, and dropout layers, as illustrated in Fig. 3(a).

In detail, our proposed model consists of four encoding function blocks, each utilizing three repeated convolutional layers with a kernel size of 3x3, followed by a ReLU activation function. Each block also incorporates a 2x2 max-pooling layer with strides of 2x2 for down-sampling. With each encoding step, the number of feature channels doubles. The first encoder block processes the input layer with a size of 256 × 256 × 3, producing two outputs: p1 with dimensions of 128 × 128 × 32 and c1 with dimensions of 256 × 256 × 32, enhanced by a max-pooling layer and a dropout layer. The second encoder block further down samples the features, producing p2 with dimensions of 64 × 64 × 64 and c2 with dimensions of 128 × 128 × 64, again including max-pooling and dropout layers. Similarly, the third block outputs p3 with dimensions of 32 × 32 × 128 and c3 with dimensions of 64 × 64 × 128, while the fourth block generates p4 with dimensions of 16 × 16 × 256 and c4 with dimensions of 32 × 32 × 256, both including max-pooling and dropout layers.

After the encoding process, the central encoding layer combines and processes the features, resulting in an output of size 16× 16 × 512. This serves as the input to the decoding path. The decoding process mirrors the encoding structure but operates in reverse to upscale the feature maps. The first decoding block integrates attention gate-1 and produces d1 with dimensions of 32 × 32 × 256. The second decoding block, connected via attention gate-2, outputs d2 with dimensions of 64 × 64 × 128. The third decoding block, linked to attention gate-3, produces d3 with dimensions of 128 × 128 × 64, and the final decoding block, connected through attention gate-4, generates d4 with dimensions of 256 × 256 × 32. The model concludes with an output layer of dimensions 256 × 256 × 1, representing the final single-channel feature map. The key hyperparameters of the proposed extended customized U-Net model play a crucial role in optimizing performance. The model utilizes ReLU and Softmax activation functions to introduce non-linearity and facilitate multi-class classification. A dropout rate of 0.5 helps prevent overfitting, while a learning rate of 0.0001 ensures stable convergence. The model is trained with a batch size of 32 over 25 epochs using the Adam optimizer for efficient weight updates. Additionally, the input image size is set to 256 × 256 × 3, ensuring consistent processing of skin cancer images.

3.1 Dataset and preprocessing results

We have collected publicly available skin cancer images with their mask, and we customized the dataset corresponding in a serious alignment with each image with its mask index. A sample image with its mask has been shown in Fig. 4. In this paper, we used two datasets, the first dataset contained the images from both publicly available datasets named as DermIS and Dermaquest (Wen et al., 2022). We have considered a second dataset named ISIC-2016 (Cassidy et al., 2022). In our major dataset, we aligned each image with its respective index in an ascending order with its respective mask. Before training compilation, we processed our image data sample, resizing the image pixel size (256 × 256) with a 3-channel feature colormap like RGB, grayscale etc. The data sample details for train test and validation have been described in a tabular format in Table 2. The dataset consists of melanoma and nonmelanoma are two categorical skin cancer images sampled with each of their masks.

Close

Fig. 4.

Some images and each image’s mask of melanoma and nonmelanoma data sample of DermIS and DermQuest Dataset. DermIS: Dermatology image search, DermQuest: Dermatology quest.

Export to PPT

3.2 Classification results

The results demonstrate outstanding performance with a 97.99% accuracy on a custom-collected image set from both DermIS and DermQuest datasets, reflecting the model’s strong ability to accurately classify and segment skin lesions. Additionally, the model achieved 91.75% accuracy on the ISIC-2016 dataset, further validating its robustness across various dermatological datasets. The graphical analysis of loss and accuracy graphs shows steady improvement during training, indicating the model’s efficient learning process. These results highlight the effectiveness of the proposed deep learning approach, enhanced by attention mechanisms and advanced encoding-decoding techniques, in achieving high precision in skin cancer detection.

3.3 Performance metrics and evaluation

In the context of evaluating the performance of U-net architecture, three key matrices are considered as Accuracy, loss, and IoU (Intersection over Union) over train and validation dataset to analyze the model prediction performance over the test case sample and the differences between train and test performance comparison. The accuracy metrics measure the proportion of correct predictions according to the model performance evaluation where the loss measures how well the model prediction matches the actual labels. In the scenario of IoU (Intersection over Union), which alternate name the Jaccard index (Bouchard et al., 2013), evaluates overlapping between the predicted segmentation and the ground truth. Between the true map, it calculated the area of overlap that is divided by union area. In Fig. 5(a), we plotted our extended deep U-net model performance over DermIS and DermQuest Dataset discussed in subsection 3.1 and in Fig. 5(b) over ISIC(2016) dataset in a graphical representation where the y axis represents the values of performance in percentage and the X-axis represent the number of epoch in every subfigure as follow: Subfigure-( a) represents the loss function, Subfigure-( b) represent the accuracy curve and Subfigure-( c) represent the Intersection over Union (IoU) curve, collectively demonstrate the model’s strong performance, showcasing effective learning, high classification accuracy, and precise segmentation capability.

Close

Fig. 5.

Graphical analysis of the proposed model’s performance across different datasets. (a) Illustrates the performance metrics on the ISIC (2016) dataset, while (b) Depicts the results on the DermIS and DermQuest datasets. These visual representations provide a comparative evaluation, highlighting the model’s effectiveness in skin cancer classification across diverse datasets. DermIS: Dermatology image search, DermQuest: Dermatology quest, ISIC: International skin imaging collaboration.

Export to PPT

3.4 Comparison with existing methods

Table 3 presents a comparative analysis of our proposed extended U-Net model against existing methods for skin cancer classification and segmentation. While previous models such as SegNet (Badrinarayanan et al., 2017), U-Net variants (Turukmane et al., 2023; Anand et al., 2023), and MFO-Fuzzy U-Net (Bindhu and Thanammal, 2023) have achieved high accuracy, our extended U-Net model outperforms them with an accuracy of 97.99%. Compared to Anand et al. (Anand et al., 2023) (97.96%) and Bindhu et al.(Bindhu and Thanammal, 2023) (97.57%), our model achieves a slight but significant improvement, demonstrating its effectiveness. The incorporation of advanced encoding-decoding blocks and attention-gating mechanisms enhances feature extraction and localization, leading to more precise segmentation and classification. These improvements validate the robustness and superior performance of our proposed approach over existing methods.

4.1 Analysis of experimental results

Guided grad-cam (GCAM) and GCAM++ are the visualization and image segmentation techniques used to understand and interpret the process of decision-making. A feasible system of our proposed extended U-net segmentation model provided Original mask, Predicted mask, and Grad-CAM visualization in Fig. 6(a) over DermIS and DermQuest (Skin cancer images with Masks) data samples and in Fig. 6(b) over ISIC(2016) data samples. Fig. 6(c) illustrates the original, predicted, and processed masks for melanoma and NMSC images, highlighting the segmentation effectiveness on the DermIS and DermQuest datasets. Fig. 6(d) focuses specifically on melanoma skin cancer, showcasing the original, predicted, and processed masks from the ISIC (2016) dataset, demonstrating the model’s precision in identifying affected regions. In the context of image classification tasks, these techniques are used to identify a certain area to capture the highlighted area of an input image sample which contributes the most to the network prediction. After computation, the gradient, as claimed by the predicted level score Grad-CAM, generates a heat map with respect to the feature maps of the last convolutional layer, which contain the same input height and width but might be, with different channel numbers. In Table 3, we compare our extended U-net model with other existing models in detail, those used in the medical section on the subject of image segmentation and visualization.

Close

Fig. 6.

Grad-CAM visual explanations of predicted data samples, highlighting the model’s focus areas for skin cancer classification. (a) Illustrates results on our customized DermIS and DermQuest datasets, while (b) Showcases predictions on the ISIC (2016) dataset. (c) and (d) Compare the original, predicted, and processed masks for melanoma and nonmelanoma classifications. Specifically, (c) Visualizes these masks for the DermIS and DermQuest datasets, whereas (d) Focuses on melanoma predictions within the ISIC (2016) dataset. These visualizations provide insights into model interpretability and segmentation accuracy. Grad-CAM: Gradient-weighted Class Activation Mapping, DermIS: Dermatology Image Search, DermQuest: Dermatology Quest, ISIC (2016): International Skin Imaging Collaboration (2016).

Export to PPT

4.2 Challenges and future works

Our proposed extended U-Net architecture achieves an impressive 97.99% accuracy in classifying melanoma and non-melanoma skin cancer, demonstrating its robustness. The integration of Attention Gates enhances feature selection by focusing on critical lesion regions, improving segmentation precision. Skip connections help retain spatial information, while image augmentation boosts model generalization. However, despite its high accuracy, the model has limitations. It may struggle with highly imbalanced datasets, leading to potential bias toward dominant classes. Additionally, the computational cost is higher due to complex attention mechanisms and deep encoder-decoder pathways. Future improvements can focus on optimizing inference speed and handling rare, ambiguous lesion cases more effectively. Future advancements in skin cancer detection should address key challenges such as dataset biases, model generalizability, and real-world applicability. Many datasets contain imbalanced distributions, where non-melanoma cases outnumber melanoma cases, potentially leading to biased predictions. Enhancing data diversity with multi-source data sets can improve robustness. Additionally, while our extended U-Net with Attention Gates achieves 97.99 % accuracy, its performance may vary across different demographics, imaging conditions, and skin types. Deploying the model in real-world clinical settings requires careful validation, handling of noisy images, and integration with dermatological workflows. Future research should explore lightweight architectures for faster inference, domain adaptation techniques for broader generalization, and AI-driven decision support systems to assist dermatologists in early and accurate skin cancer diagnosis.