Novel bioluminescent oceanic optimization with explainable AI for accurate skin cancer diagnosis and enhanced dermatologist support

2.1 Machine learning and deep learning techniques

The parallel convolutional spiking neural network (PCSN-Net) (Kumar, K. A. and Vanmathi, C. 2025) is used for efficient skin cancer diagnosis. Achieving 95.7% accuracy, 94.7% sensitivity, and 92.6% specificity, PCSN-Net demonstrated strong diagnostic performance. Fully convolutional deep supervised convolutional neural network (FCDS-CNN) model (Nawaz, K. et al. 2025) optimized for early melanoma detection. It reached an average accuracy of 96 % in seven classes by using class weighting and data augmentation, which improved the work with imbalanced datasets. A convolutional neural network is prioritized, and the region-based convolutional neural network (R-CNN) algorithm is employed to find minor changes in skin lesions (Kavitha, C. et al. 2024). A CNN-based classification system (Vibha, T. R. et al. 2025, Reddy, A. S. and MP, G. 2025) have suggested a dual-component model based on the proposed segmentation using optimized Mask R-CNN and the GEMLO model, where the optimal feature selection is based on the framework. Hybrid architecture uses InceptionV3 and DenseNet121 as feature extractors, and then fuses the resultant representations with weighted sum rule is based on that (Akter, M. et al. 2025). This was improved by adopted a heterogeneous ensemble learning model that was based on stacked support vectors machines (Das, A. and Mohanty, M. N. 2025). A multimodal model, which combines dermoscopic images with synthetically generated clinical notes, with the help of GPT-4-turbo and the Align Before Fuse (ALBEF) framework (Chakkarapani, V. et al. 2025). A lightweight melanoma screening approach which combined Gaussian filtering, deep learning–based nonlinear data processing (DLNDP) feature extractor, and hybrid randomized adaptive neuro-fuzzy inference system (RANFyIS) detection model (Sargunan, I. et al. 2025), achieving 90.42% accuracy with a high true positive rate of 92.46%. The same architecture and confirmed reliability with consistent metrics, reinforcing potential for non-invasive and rapid melanoma screening in clinical settings (Swapno, S. M. R. et al. 2025).

2.2 Explainable AI in skin disease

Meta-explainable privacy-preserving federated learning (Meta-XPFL) framework is used for skin cancer diagnosis using IoMT images. XAI provided interpretability, addressing clinical usability (Serhani, M. A. et al. 2025). IARS SegNet, a lesion segmentation network based on SegNet with residual convolution and global attention (Narayanan, V. S. et al. 2024). YOLOv8 is used for Region of interest (ROI) detection, enabling better lesion localization and focus before classification (AbuAlkebash, H. et al. 2025). A hybrid model integrates ViT, a custom CNN, and Xception, supported by Grad-CAM for visual explanation (Ali, A. et al. 2025). However, reliance solely on visual features restricted the inclusion of other modalities specific knowledge. Synthetic minority over-sampling technique (SMOTE) is used for class balancing and used CNN and ResNet50 for feature extraction, along with LIME for local interpretability (Gupta, R. K., 2025). To fill this gap, a Trustworthiness Index for XAI using EfficientNet-B0 (Ieracitano, C. et al. 2025). Modified UNet-inspired encoder-decoder architecture is preserved spatial features for precise melanoma segmentation (Kibriya, H. et al. 2025). To bridged this gap by assessing XAI through eye-tracking in clinical setting (Chanda, T. et al. 2025). The model transparency by comparing CNN, Xception, EfficientNet, VGG, and ResNet using XAI techniques like LIME and Grad-CAM across skin cancer, brain tumor, and pneumonia detection (Garg, P. et al. 2025). The concept of explainable diagnosis by Mpox-XDE, a hybrid deep model combining Xception, DenseNet201, and EfficientNetB7 (Kumar Saha, D. et al. 2025) for multi-disease classification, monkey pox, and skin conditions. With 98.8% accuracy of Grad-CAM for visualizing predictions, the model illustrated the effectiveness of XAI not only in diagnosis.

Problem identification: From the comprehensive review (Table 1), critical research gaps in the current landscape of ML, DL, and XAI for skin disease diagnosis have been identified. These include poor performance on unbalanced datasets, lack of precise lesion boundary detection, limited clinical usability due to absence of interpretability mechanisms, restricted adaptability from single-model dependencies, weak generalization across datasets, underutilization of handcrafted domain-specific features, dependency on image-only analysis without contextual clinical data, high computational complexity unsuitable for real-time applications, and lack of transparent, explainable decision-making. In the domain of XAI-driven models either fail to deliver lesion-level segmentation, suffer from high model complexity, and lack comprehensive multi-modal fusion for richer diagnostics. Weak spatial segmentation consistency, absence of disease-specific differentiation, and scalability issues remain a problem, hence restricting its practical application in clinical settings. To fill these complex gaps, the current study offers an XAI architecture built around an optimal encoder-decoder model, a combination of the ResNet50 and EfficientNet-B4 to extract all possible features, squeeze attention network (SAttNet) to selectively locate lesions, and a hybrid decoder that is based on the U-Net and SegNet to ensure the accurate division of parts and structural arrangement. Bioluminescent oceanic optimization (BOO) algorithm is used to ensure successful feature selection whereas SHAP, LIME, and Grad-cAM are added to ensure post-hoc explainability. Such a combined model will not only enhance the accuracy of diagnosis in a wide variety of datasets, but it also leads to clinical trust due to the transparent and interpretable decision-making, which is solution to dermatologist-aided skin cancer diagnosis.

3.1 Data preparation and augmentation

This paper used composite of publicly accessible dermoscopic image datasets, namely ISIC last year (2018), ISIC last year (2019), ISIC last year (2020), and HAM10000 skin image datasets (Purni, J. T. and Vedhapriyavadhana, R. 2024) for the development of skin lesion classifier. The HAM10000 data set consists of 10,015 images, classified into 7 diagnostic classes: dermatofibroma (DF), melanocytic nevi (NV), melanoma (MEL), benign keratosis-like lesions (BKL), basal cell carcinoma (BCC), actinic keratoses (AKIEC), and vascular lesions (VASC). Dermoscopic images are shown in Fig. 2 for each class as a representative image. Geometric transformations, such as horizontal and vertical flips, 90 degrees and 180 degrees rotations, etc., were used as well as quality-based augmentations, such as Gaussian blur and additive Gaussian noise, as data-augmentation techniques. These operations simulate real life variations and distortions, thus increasing the robustness of the model to various inputs. All original and augmented images were processed using an improved Canny edge detector (ICED) module to highlight the lesion boundary features. The preprocessing pipeline was set up to be compatible and uniform across all input images. In order to get roughly the same input dimensions and reduce computation burden, images were automatically down sampled to 32*32 pixels. Pixel values were then normalized either in the [0, 1] or [- 1, 1] range, depending on the normalization needs of the model being trained; the normalization made training easier and increased the rate of convergence. Depending on the classification strategy used, diagnostic class labels were expressed either as one-hot vectors, or as integer encodings. After augmentation and integration of ICED, the entire dataset was converted to NumPy arrays, allowing for efficient manipulation, storage, and ingestion by the model. Table 2 shows a summary of the distribution of the samples belonging to different classes in the datasets before and after augmentation.

Close

Fig. 2.

Sample dermoscopic images of each class from (a) ISIC 2018 [MEL, NV, BCC, AKIEC, BKL, DF, VASC], (b) ISIC 2019 [MEL, NV, BCC, AK, BKL, DF, VASC, SCC], (c) ISIC 2020 [Benign, Malignant], and (d) HAM10000 [AKIEC, BCC, BKL, DF, MEL, NV, VASC].

Export to PPT

3.2 Encoder

The encoder, in this case, will be the central element of the automated feature recognition of dermoscopic images. The encoder is also based on the idea of capturing a hierarchy of features (starting with low-level visual features, like edges, contours, and colour gradients, and then high-level semantic features, such as asymmetry, irregular pigmentation, and border discontinuities, which are clinically relevant features of malignancy). In order to obtain strong and various feature representation, the encoder of the proposed framework combines two strong pre-trained convolutional neural network backbones: ResNet50 and EfficientNet-B4. ResNet50 (Firasari, E. et al. 2024) is a 50 layer deep residual network which uses residual learning by using skip connections to successfully overcome the vanishing-gradient problem. EfficientNet-B4 (Dinitra, N. T. et al. 2024) implements a compound-scaling method that is most optimal in terms of network depth, width and input resolution to establish high accuracy and to discriminate between lesion borders, globules, dots, and streak as well as lesions, which are important clinical indicators in skin-cancer.

3.3 Decoder

The decoder plays a vital role in the process of obtaining valid outputs of the encoded feature representations by performing two main functions (i) accurate lesion segmentation and (ii) multi-class skin cancer classification. To achieve the objectives successfully, dedicated decoder architecture is utilized, combining two fully-developed models, including UNet and SegNet models, which, despite being initially defined with a particular focus on biomedical image segmentation, are extended in this case, to serve the purpose of classification. UNet (Kibriya, H. et al. 2025) is an image segmentation network which is designed to use symmetric encoder-decoder architecture, with skip connections to preserve spatial detail. In this context, UNet enables the decoder to retain critical boundary and texture information, which is essential for identifying the irregular shapes and edges of skin lesions. SegNet (Kumar, K. A. and Vanmathi, C. 2024) contributes to structural refinement in the decoding process which enhances semantic segmentation by storing max-pooling indices from the encoder and reusing them during upsampling. SegNet supports the classification branch and helps maintain accurate lesion shapes in the final output.

3.4 Feature selection

The BOO is bio-inspired algorithm used for the skin lesion classification using dermoscopic images (Fig. 3). Drawing inspiration from the bioluminescent behavior of marine organisms such as plankton and jellyfish, which emit light signals for communication, navigation, and survival in the dark ocean depths, BOO mimics these adaptive strategies (Qin, X. et al. 2024) to navigate the complex, high-dimensional feature space derived from dermoscopic image data. The mechanism of attraction and repulsion describes how the agents tend to gravitate towards more informative sets of features -through brighter agents, and also avoids suboptimal sets of features, and thus equilibrium between exploration and exploitation is maintained (Halbers, L. P. et al. 2024). Unlike other traditional paradigms like the GA, PSO, and ACO (Wu, Y. et al. 2025), which often undergo premature convergence and early exploration, BOO uses light based signaling to guide agents to optimal informative feature subsets. The agents (M) are represented either by a binary vector (mY) or by a real-valued one (mY, P) representing the selection status of the features (mY C), and the initialization of the whole population of BOO is performed in the following way.

Close

Fig. 3.

BOO algorithm for feature selection.

Export to PPT

The initial state of the bioluminescent agents in the search space is generated using randomized initialization at the beginning of the algorithm execution, where each agent represents a candidate feature subset for the classification task.

Here, M represents the population matrix of bioluminescent agents, where the h-th agent (a candidate feature subset) and its g-th dimension correspond to a specific decision variable (feature) in the search space. Y denotes the whole numeral of agents, and d is the numeral of features (decision variables). The values are initialized using random numbers in the interval [0, 1]. Each agent thus represents potential feature combination and evaluates the objective function based on its selected feature set.

The vector F represents the objective function values corresponding to each candidate feature subset in the population. The global search strategy enhances diversity and helps avoid premature convergence.

This represents a scenario where a bioluminescent agent explores a promising region in the search space by moving toward the best-known feature subset (i.e., the current optimal solution).

where $M_h^{A1}$represents the new location of the h-th founded on the first phase, $m_{h,g}^{X1}$ is its g-th dimension, $f_h^{A1}$is its objective function worth, $R_{h,g}$are random statistics in interval [0, 1].

If the newly generated feature subset improves the objective function—such as classification accuracy or feature compactness—it replaces the previous subset for that agent, ensuring continuous enhancement of the solution quality.

Based on the $M_h^{A2}$ second phase of BOO algorithm, h-th represents the new state of the fitness, $m_{h,g}^{A2}$ its g-th measure $f_h^{A2}$ represents its objective function. Upon completion, the best feature subset identified throughout the optimization process is selected as the final optimal solution for feature selection.

3.5 Attention enhancement

Attention enhancement refers to the use of mechanisms that assist the typical to emphasis selectively on the most important areas of the dermoscopic imageries, particularly the areas around lesion boundaries that are critical for accurate diagnosis. The SAttNet is used to refine the article maps generate by the strength system by emphasizing diagnostically relevant information while suppressing irrelevant background noise. SAttNet (Ocal, H. 2024) works by first performing a squeeze operation, such as comprehensive regular pooling, to aggregate spatial info into a compact descriptor. As shown in Fig. 4, SAttNet captures rich multi-scale longitudinal structures and efficiently begin long-term $F_p\in R^{D\times I\times Z}$ relationships between channel foci at different scales. D stands for the numeral of networks in an input article map, whereas I and Z stand for the feature map’s size. To improve the volume’s depiction of vascular characteristics, the feature maps of every branch were included.

Close

Fig. 4.

Construction of SAttNet.

Export to PPT

Based on the bifurcated input paradigm, $F_t^y$the computation that can be performed on each branch of the transition (T_eff) is as the following:

The non-linear vector quantization $f_{conv}$ belongs to the attention mechanism with the objective function K^(a), and the kernel size is fixed by convolutional vector. The fitness optimization $f_{conv}$ performs by the maximum threshold vector follows the optimal solution allows to reach convergence matrix (J^(a)). The joint cluster vector allows fixing the convolutional layer in the Attention mechanism.

The weight part consists of two stages; compression and excitation. The feature map’s $J_{dis}$ spatial dimension is compressed using the average aggregation. The objective function of each candidates follows the maximum rule-set theory to fix the optimal solution.

Two repeated fully linked deposits were utilized to learn and give masses to every frequency during the stimulus phase, reflecting the significance belongs to distance between threshold vectors (T_diff).

The optimal vector function is used to formulate the maximum threshold J_d function by using the ReLU. The control vector for convolutional layer is formulated by the sigmoid function Att^(a) with the difference between minimum Z₁ and maximum Z₂ threshold vectors.

The channel consideration is considered a viable aspect in the segmented function $F_t^{\left(a\right)}$, which allows the linear vector formation to the attention module.

The control values for fixed attention vector are compute through the maximum threshold.

The collections of oral records, catalogued in terms of frequency, are thrown away to come up with the final cartographic image as demonstrated by the resulting tracks:

In this context, the Multi-Scale Attention (MSA) component $F_{out}\in R^{D\times I\times Z}$ is generated by a multiplicity of calibrated processes. The loss utility $l_{Dice}$ and $l_{cldice}$ can be expressed as the weighted aggregation of the relevant metrics and loss functions, as explicated in the further analyses.

The vessel subdivision $l_{Dice}$ is accurate or not depends on a confluence of disciplinary views $l_{cldice}$ which affect, in their turn, the richness of $l_{Dice}$ the receptacle structure which lies behind them. In this regard, the predetermined mindset about a balance problem that occurs due to the presence of various influencing factors is an issue $l_{cldice}$. The rational feature separation predictive model and correspondence through the Dice loss function are customized as the following.

The smooth quantization vector (L_SQV) is optimal function to fix the t$v_g$hreshold by the occurrence mutual vectors.

The random solution follows the square low to fix the quality measure functions of decision model which fix the objective function of linear topological vector.

3.6 Explainability and interpretability

An XAI is incorporated to help alleviate this problem by allowing the model’s predictions to be interpreted both locally and globally.

• The gradient-weighted class activation mapping (Grad-CAM) (Na, A. 2024) is used to generate visual explanations by produces heatmap over input images. The geographical areas that the model concentrates on while generating categorization decisions are highlighted in these Heatmaps.

• The SHAP (Hu, B. et al. 2024) offers a quantitative interpretation by assigning each feature a contribution value based on influence on the model’s output. SHAP provide both global insights and local insights, making it highly valuable for model debugging and validation.

• The local interpretable model-agnostic explanation (LIME) (Mahmud, M. Z. et al. 2024) complements these methods by offering instance-specific explanations. For each input case—such as a patient’s dermoscopic image—LIME identifies and visualizes which specific regions most influenced the model’s classification.

4.1 Impact of feature selection in skin cancer diagnosis

The experimental analysis presented in Fig. 5 for the ISIC Archive dataset evaluates the performance of six feature selection algorithms—GA, PSO, GWO, WOA, ACO, and BOO—across three key metrics: feature selection time, convergence speed, and generalization. BOO is also shown to be significantly better than other algorithms in all the metrics evaluated in both figures. These comparative findings provide conclusive evidence that burrow-based owl optimization (BOOW) is the most efficient feature-selection tool in the set of tools considered, with the shortest processing time, the fastest convergence, and the best generalization performance, which makes it appropriate for the target of detecting skin-lesions in dermatological data. The findings indicate that EfficientNet-B4+UNet+BOO+SAttNet model has the best performance in the skin cancer diagnosis, offering high classification accuracy, strong generalization, and decent class-level discrimination in all lesion types in the ISIC 2018 dataset. The efficiency of the proposed BOO algorithm can be clearly seen in the comparative analysis of the computational complexity of various feature-selection algorithms, which is available in Table 3. BOO has the shortest average time of computation, that is equal to 64.19s with just 60 iterations, indicating a 45.81% reduction with respect to GA, 38.01% reduction with respect to particle swarm optimization (PSO), 29.7% reduction with respect to GWO, 26.4% reduction with respect to whale optimization algorithm (WOA), and 49.25% reduction with respect to ant colony optimization (ACO). It is an improvement that highlights the better optimization and convergence, and capabilities of BOO. Its computational efficiency is also supported by the fact that the corresponding theoretical complexity of the O(N × D) is lower than its competitors, including GA and ACO, have higher complexities of O(N x D x G) and O(N 2 x D) respectively, which in turn increases their run time. The success of light-intensity-based exploration mechanism enabling adaptability of BOO has largely contributed to its success in gaining performance in terms of exploitation and exploration, reducing unnecessary evaluations, and convergence speed.

Close

Fig. 5.

Results of feature selection algorithms for ISIC Archive dataset (a) Feature selection time, (b) Convergence Speed, (c) Generalization (Overfitting Gap %).

Export to PPT

4.2 Performance analysis of ISIC 2018 dataset

The experimental outcomes shown in Fig. 6 give a detailed comparison between a number of advanced model of learning under four important performance measures; Accuracy, Precision, Recall, and F-measure- within the framework of skin lesion classification. The architecture that included EfficientNet-B4, UNet, BOO, and SAttNet provided the best accuracy results of all lesion types, where the result was as high as 99.817% on nevus (NV), 99.786% on benign keratinocyte lesions (BKL), and consistently high results on the other classes (MEL, NV, BCC, AKIEC, BKL, DF, VASC). The setup that used EfficientNet-B4, SegNet, BOO, and SAttNet yielded slightly worse but still very competitive results, which demonstrates that both UNet and SegNet frameworks can be effectively used with the help of BOO and SAttNet modules. On precision, once again EfficientNetB4UNetBOO SAttNet setup performed better with top scores in NV of 99.803% and BKL of 99.761% with low false positives. The same pattern was observed in the case of recall values as the same model obtained 99.791% for NV, 99.737% in basal cell carcinoma (BCC), and the lowest inter-class deviation. The F- measure that weighs precision and recall was also the highest at this composition in all classes reaching 99.796% in the case of NV and 99.752% in the case of BKL hence indicating an optimal trade-off in classification. From Figs. 7 and 8 shows that the BOO optimization algorithm significantly improved performance across all metrics. When comparing models with and without BOO, the performance uplift is clearly visible. For example, ResNet50+UNet+SAttNet had a maximum accuracy of 99.187%, while with BOO, the improved version (ResNet50+UNet+BOO+SAttNet) achieved up to 99.583%.

Close

Fig. 6.

Results of proposed models for ISIC 2018 dataset (a) Accuracy, (b) Precision, (c) Recall, and (d) F-measure.

Export to PPT

Close

Fig. 7.

Training and Validation performance of the proposed model for skin cancer diagnosis on the ISIC 2018 dataset (a) Accuracy (b) Loss.

Export to PPT

Close

Fig. 8.

Confusion matrix for ISIC 2018 dataset.

Export to PPT

4.3 Performance analysis of the ISIC 2019 dataset

The quantitative analysis shown in Fig. 9 evaluates the presentation of several deep learning configurations over eight skin lesion classes—MEL, NV, BCC, AK, BKL, DF, VASC, and SCC—using the ISIC 2019 dataset. The most successful classification accuracy was observed with the EfficientNet-B4+UNet+BOO+SAttNet architecture, with the performance metrics ranging between 99.589% seborrheic keratosis (SCC) to 99.816% nevi (NV), and thus proving to have a strong discriminative performance across regularly and uncommonly observed lesion subtypes. Fig 9-11 of the experimental analyses show that the EfficientNet-B4+UNet+BOO+SAttNet model is offering the best and consistent results in all the eight types of lesions in the ISIC 2019 dataset. Both the integration of the BOO strategy can help optimize and converge more efficiently. The SAttNet module can complement spatial feature awareness and make the suggested architecture especially suitable to be implemented as a practical clinical decision-support device in skin-cancer diagnostics.

Close

Fig. 9.

Results of proposed models for ISIC 2019 dataset (a) Accuracy, (b) Precision, (c) Recall, and (d) F-measure.

Export to PPT

Close

Fig. 10.

Training and Validation performance of the proposed model for skin cancer diagnosis on the ISIC 2019 dataset (a) Accuracy (b) Loss.

Export to PPT

Close

Fig. 11.

Confusion matrix for ISIC 2019 dataset.

Export to PPT

4.4 Performance analysis of the ISIC 2020 dataset

The results of the comparison of the performance of a few hybrid deep-learning models on the ISIC 2020 data on skin-cancer classification are shown in Fig. 12. The EfficientNet-B4+UNet+BOO+SAttNet setup was the most successful in its comparison, and had the highest rates of accuracy of 99.742% on the Benign and Malignant classes, respectively. The combination of the BOO and SAttNet was important in the refined feature selection and spatial attention respectively, led to an increased confidence in lesion recognition and classification respectively. As it can be seen in Figs. 12-14, the EfficientNet-B4+UNet+BOO+SAttNet model provides high and stable results on all the eight lesion types in the ISIC 2020 dataset. The BOO component helps to improve optimization and convergence, and the SAttNet module is used to improve the awareness of the spatial features, but the given architecture is extremely appropriate to the practical clinical decision-support in skin-cancer diagnostics.

Close

Fig. 12.

Results of proposed models for ISIC 2020 dataset (a) Accuracy, (b) Precision, (c) Recall, and (d) F-measure.

Export to PPT

Close

Fig. 13.

Training and Validation performance of the proposed model for skin cancer diagnosis on the ISIC 2020 dataset (a) Accuracy (b) Loss.

Export to PPT

Close

Fig. 14.

Confusion matrix for ISIC 2020 dataset.

Export to PPT

4.5 Performance analysis of HAM10000 dataset

As shown in Fig. 15, the experimental assessment of HAM10000 dataset is a detailed, class-related analysis of the suggested models in the seven skin-lesion categories actinic keratoses and intraepithelial carcinoma (AKIEC), Basal cell carcinoma (BCC), Benign keratosis-like lesions (BKL), Dermatofibroma (DF), Melanoma (MEL), Melanocytic nevi (NV), and Vascular lesions efficientNet-B4 – Efficient network (Baseline-4 variant)+U-shaped network (UNet)+Bioluminescent ocean optimization (BOO)+Squeeze attention network (SAttNet) model supported by AKIEC, NV, and VASC resulted in an accuracy of 99.687%, 99.831%, and 99.707% which indicated its ability to classify both uncommon and common type of lesions with an exceptionally high degree of accuracy. On precision, BCC and NV had a 99.742% and 99.799%, respectively, which is excellent; it represents an excellent proportion of truly identifying a positive result with a low number of false positives. The strong sensitivity of the model in identifying even the most visually ambiguous classes was supported by the recall values of 99.547% of MEL and 99.724% of BCC. The F-measure scores of 99.733% BCC and 99.783% NV supported the consistency and validity of the model in the accuracy and consistency in complex cases with overlapping features. Figs. 15-17 also indicate that EfficientNet-B4+UNet+BOO+SAttNet model gives better and stable results in all the eight lesion classes of HAM10000 dataset. BOO addition improves optimization and convergence, and the SAttNet addition improves spatial feature knowledge, which is why the proposed model is extremely applicable in real-life clinical decision-support in skin-cancer diagnosis.

Close

Fig. 15.

Class-wise result analysis of proposed models for HAM10000 dataset (a) Accuracy, (b) Precision, (c) Recall, and (d) F-measure.

Export to PPT

Close

Fig. 16.

Training and Validation performance of proposed model for skin cancer diagnosis on HAM10000 dataset (a) Accuracy (b) Loss.

Export to PPT

Close

Fig. 17.

Confusion matrix for HAM10000 dataset.

Export to PPT