The DMR-IR database is a publicly accessible repository containing multimodal breast examination data, including infrared thermography, digital mammography, and clinical records. For this study, only the frontal thermal images of 130 patients were used, acquired under the First Static Protocol to ensure standardized conditions. Patients included both healthy controls and individuals with benign and malignant breast lesions, thereby introducing variability essential for robust model evaluation. The infrared images were captured using a FLIR SC620 camera, with a sensitivity of <0.04 °C and a temperature range of 40 °C–500 °C, at a resolution of 640 × 480 pixels. To minimize external variability, all acquisitions followed a controlled clinical protocol, where patients were acclimatized in a room maintained at 20 °C–23 °C for 15 min before imaging. Manual annotations were performed by the study authors to generate ground-truth masks, with cross-verification among annotators to reduce bias. While not performed by certified radiologists, this procedure was explicitly designed for experimental, non-clinical purposes. The use of the DMR-IR database ensures transparency and reproducibility, as the dataset is publicly available online (http://visual.ic.uff.br/dmi), allowing independent research.

In this study, the dataset utilized for experimentation was obtained from an accessible online database. The main objective of this research revolves around the comparative analysis of three U-Net deep neural network variants. Therefore, leveraging a pre-existing dataset rather than dedicating resources to the creation of a new one is a strategic decision which allowed this study to focus on the main objective.

To improve breast region segmentation in thermal images, this study evaluated three distinct deep learning architectures. Each of these models is recognized for their image segmentation ability.

The training and evaluation of the U-Net, U-Net with Spatial Attention, and U-Net++ models were conducted based on the flowchart of the algorithm as illustrated in Figure 6.

The algorithm begins by loading grayscale images and its corresponding true masks from specified directories, resizes, and normalizes their pixel values. To augment the dataset, the images and masks are subjected to various transformations as explained in Section 4.1.3, thereby introducing variability into the training data. A stratified split approach was adopted to split the data into training and validation sets, with 20% of the augmented dataset reserved for validation. This is specifically for experimental comparison between models. This allocation, while not subjected to active experimentation, was designed to ensure a balanced representation of diverse classes in both training and validation sets. The model is then initialized based on the specific model type (U-Net, U-Net with Spatial Attention, or U-Net++). The model is compiled using the binary cross-entropy loss function and accuracy metric, crucial for measuring segmentation precision.

Five different optimizers were evaluated comparatively: ADAM, NADAM, RMSPROP, SGDM, and ADADELTA. The evaluation of the specific optimizers was based on their widespread usage and documented effectiveness in various deep learning applications, particularly in image segmentation tasks. The mathematical equations that describe how the optimizers update the model weights during training are as follows.

Quantitative analysis was conducted among the three deep learning models: U-Net, U-Net with Spatial Attention, and U-Net++. The flowchart in Figure 7 outlines the systematic process of conducting k-fold cross-validation analysis for evaluating the three pre-trained segmentation models.

The evaluation begins by loading the pre-trained model. Test images and their corresponding masks are then loaded, and the dataset is divided into subsets for k-fold cross-validation. Within each fold, the data is further split into training and validation sets. For each validation image, it is resized to match the model’s input shape and then preprocessed. The model predicts masks for these images, which are converted into binary format. Various evaluation metrics are calculated, and the results, including the original image, true mask, and predicted mask, are visualized for inspection. After evaluating all validation images in a fold, the metrics are aggregated. Mean and standard deviation of the metrics are computed across all folds. The results are presented in Section 5.2. These provide a comprehensive overview of the model’s overall performance and its consistency across different subsets of the dataset.

Grad-CAM calculates the importance weights by performing global average pooling on the gradients of the target class with respect to the feature maps. This is mathematically represented by Equation 25 as: α k c = 1 Z ∑ i ∑ j ∂ Y c ∂ A i j k (25)

Where:

α k c is the importance weight for the k − th feature map in the c − th class,

Z is the normalization factor,

Y c is the final prediction score for class c ,

A i j k is the activation in the k − th feature map at position ( i , j ).

The weighted sum of feature maps is computed using Equation 26 to obtain the heatmap, denoted as L Grad ‐ CAM c : L Grad ‐ CAM c = ReLU ∑ k α k c A k (26)

Where:

A k represents the k -th feature map.

The overlay operation in Equation 27 involves combining the Grad-CAM heatmap L Grad ‐ CAM c with the original image I : Resultant Image = Heatmap Weight × L Grad ‐ CAM c + 1 − Heatmap Weight × I (27)

Where:

The Heatmap Weight determines the intensity of the heatmap overlay.

The Grad-CAM heatmap visualization offers valuable insights into the interpretability of the model. It aids in understanding which regions of the input images are pivotal for the model’s predictions, thereby contributing to the overall assessment of the model’s performance in thermography-based breast region segmentation. The results of the qualitative analysis are presented in Section 5.3.

To determine the optimal optimizer for training the deep learning models, a thorough comparative evaluation was conducted using five different optimizers: ADAM, NADAM, RMSPROP, SGDM, and ADADELTA. The evaluation focused on three key metrics: final loss, final accuracy, and training time. These metrics provide insight into the efficacy and efficiency of each optimizer in training the segmentation models. The final loss measures how well the model fits the training data, the final accuracy indicates the proportion of training data correctly classified by the model, and the training time reflects the optimizer’s computational efficiency. The results of this comparative evaluation are presented in Table 1–3. Which are also graphically represented in Figures 9–11.

The findings presented demonstrate significant variations in the efficacy of different optimization algorithms when applied to the three distinct deep learning models, U-Net, U-Net with Spatial Attention, and U-Net++. ADAM emerges as the preeminent optimizer for these models, consistently yielding the most favorable outcomes in terms of reduced loss values and heightened accuracy scores. The U-Net model, when trained using the ADAM optimizer, demonstrated notable performance with a final loss of 0.0357 and an accuracy of 0.9637. These results are highly competitive when compared to the others. In relation to the duration of the training, both the U-Net and U-Net with Spatial Attention architectures exhibit a notable level of efficiency, demonstrating comparable or reduced training times when compared to the U-Net++ model, regardless of the optimizer employed. The U-Net++ model exhibits a consistently longer training duration, which can be attributed to its intricate architectural design. The NADAM and RMSPROP optimizers exhibit comparable performance, albeit with marginally elevated loss values and diminished accuracy scores in comparison to the ADAM optimizer. In contrast, SGDM demonstrates notably elevated loss values and diminished accuracy scores across all three models, suggesting that it may not be the optimal selection for these specific models. Among the five optimizers, ADADELTA exhibits the poorest performance, characterized by significantly elevated loss values and notably decreased accuracy scores, along with relatively long training durations.

Using 20%–30% of the data for validation is a common practice in deep learning and medical image classification studies, as it provides a balance between training and validation sizes (Szegedy et al., 2016). While this proportion allows for an initial assessment of model performance, we acknowledge that the dataset size remains limited for testing the model’s generalizability to new or unseen data. Regarding the fixed 30 training epochs, this number was chosen to maintain consistency across all models and optimizers, with careful monitoring of training and validation loss to ensure convergence and prevent overfitting. Through our experiments, these settings proved suitable for all models to achieve optimal performance within the current dataset scope (Ding et al., 2022).

In conclusion, the ADAM optimizer is implemented for the training of the three segmentation models for its superior performance in this study. The training process of the models, using the ADAM optimizer, is visualized using their loss and accuracy graphs over the number of epochs. Figure 12, depict the convergence of each model throughout the training process. The training loss metric serves as a measure of the model’s ability to fit the data, whereas the accuracy metric reflects the frequency with which the model’s predictions align with the actual outcomes. Over the course of 30 epochs, the models demonstrated a progressive decline in the loss values and a steady improvement in accuracy, while guided by the ADAM optimizer.

K-fold cross-validation analysis was conducted to evaluate the three pre-trained segmentation models using 30% of the entire dataset. Detailed metrics, as outlined in Section 4.5, were meticulously examined. The evaluation results for U-Net, U-Net with Spatial Attention, and U-Net++ are presented in Tables 4–6, respectively. Each table provides a detailed breakdown of metrics such as Intersection over Union (IoU), Dice coefficient, precision, recall, sensitivity, specificity, pixel accuracy, ROC-AUC, and PR-AUC for every fold (k = 1–10).

The U-Net model exhibited robust performance with an average IoU of 0.9292 and a standard deviation of 0.0136. Notably, the Dice coefficient averaged at 0.9630, indicating a high degree of accuracy in segmentation. Precision, recall, and specificity consistently maintained their values across folds, underscoring the model’s effectiveness in classifying true positives and negatives. Pixel accuracy reached an average of 0.9703, signifying precise pixel-level segmentation. The model’s discriminative ability, as measured by ROC-AUC and PR-AUC, was substantial, averaging at 0.9680 and 0.9472, respectively.

The U-Net with spatial attention model exhibited competitive results, with an average IoU of 0.9290 and a low standard deviation of 0.0095. The Dice coefficient showed a mean value of 0.9630, underlining accurate segmentation. Precision and specificity demonstrated consistent values across folds, indicating reliable positive classification. The ROC-AUC and PR-AUC averaged 0.9704 and 0.9466, respectively, emphasizing the model’s strong discriminative ability.

The U-Net++ model showcased competitive performance, with an average IoU of 0.9251 and a standard deviation of 0.0128. The Dice coefficient reached an average of 0.9608, indicating accurate segmentation results. Precision and specificity displayed consistent values, highlighting the model’s ability to accurately classify positive samples. The ROC-AUC and PR-AUC, measuring the model’s discriminative ability, averaged 0.9665 and 0.9389, respectively.

Comparative analysis of the results from each segmentation model is conducted by observing the mean and standard deviation for the evaluation metrics which are summarized in Tables 7 and 8, which are also graphically represented in Figures 13, 14.

In terms of IoU, U-Net and U-Net with Spatial Attention demonstrate similarly high values of 0.9292 and 0.9290, respectively, with U-Net++ slightly lower at 0.9251. This metric reflects the degree of overlap between the predicted and ground truth segmentations, indicating the models’ effectiveness in capturing the target region.

The Dice Coefficient, another measure of segmentation accuracy, exhibits comparable performance among the models, with U-Net leading at 0.9630, followed closely by U-Net with Spatial Attention and U-Net++.

Precision, Recall, and Sensitivity metrics focus on different aspects of classification accuracy. U-Net consistently outperforms the other models in Precision, emphasizing its ability to minimize false positives. On the other hand, U-Net with Spatial Attention and U-Net++ show competitive performance in Recall and Sensitivity, highlighting their capacity to identify true positives.

Specificity measures the models’ ability to correctly identify true negatives, and U-Net maintains a slight advantage over the others in this regard. Pixel Accuracy, reflecting the overall accuracy of pixel-wise classification, indicates similar performance across the models.

The ROC-AUC and PR-AUC values, assessing the models’ discrimination and precision-recall trade-offs, exhibit minor variations among the models. Tables 4–6 provide the performance metrics measurement for UNet, UNet++ and UNet with spatial Attention and are discussed in Appendix A.

The standard deviations provided in Table 7 offer insights into the stability and consistency of each model’s performance across different metrics. Generally, U-Net demonstrates lower standard deviations compared to U-Net with Spatial Attention and U-Net++, suggesting more consistent results.

In summary, the evaluation metrics collectively suggest that U-Net performs competitively, demonstrating strong segmentation accuracy and consistency. U-Net with Spatial Attention and U-Net++ exhibit comparable performance, with slight variations in specific metrics. These findings contribute valuable information for selecting an appropriate model based on the desired trade-offs in thermography-based breast region segmentation.

Visual inspection of the segmentation results was conducted using the Grad-CAM heatmaps, focusing on the predicted region of interest generated by each model. Figures 15–17 display the Grad-CAM heatmaps for U-Net, U-Net with Spatial Attention, and U-Net++, respectively. The color patterns and transitions are observed from the heatmaps, providing a visual representation of how the model assigns importance to different areas in the thermal images. This visual inspection aids in understanding which regions the model identifies as crucial for predicting the presence of breast tissue, contributing to the interpretability of the model’s decision-making process.

Table 9 presents the comparative scores of Breast Region Overlap (BRO) and Noise Handling (NH) for U-Net, U-Net with Spatial Attention, and U-Net++ across 30% test images. The scores range from 1 (Poor) to 5 (Excellent) which were explained in Section 4.6.1.

Table 10 presents the comparative averaged scores of qualitative evaluations for the U-Net, U-Net with Spatial Attention, and U-Net++ models across two criteria: Breast Region Overlap and Noise Handling. A corresponding visual representation is provided in Figure 18, depicting the average scores for these models.

In terms of Breast Region Overlap, U-Net stands out with an impressive average score of 4.10, indicating a significant ability to align precisely with the breast boundaries in thermal images. On the other hand, both U-Net with Spatial Attention and U-Net++ exhibit lower average scores of 2.10 and 2.13, respectively, suggesting a diminished capability to accurately overlap with the actual breast region.

In terms of Noise Handling, U-Net excels with a high average score of 4.7, show-casing robust noise handling and concentrated activations on the breast area. In contrast, U-Net with Spatial Attention and U-Net++ struggle with noise, as indicated by their average scores of 2.53 and 1.83, respectively. These models face challenges in maintaining clarity in depicting the breast region due to noticeable noise patterns.

The overall analysis highlights U-Net’s superior performance in both Breast Region Overlap and Noise Handling compared to U-Net with Spatial Attention and U-Net++. Furthermore, U-Net with Spatial Attention and U-Net++ exhibit comparable performance, with U-Net++ showing a slight improvement.

While the evaluation metrics demonstrate strong performance across all three U-Net variants, statistical validation is essential to assess whether the observed differences are significant. A pairwise Wilcoxon signed-rank test was applied across the folds of cross-validation for IoU and Dice scores, comparing U-Net against U-Net++ and U-Net with Spatial Attention. Results indicated no statistically significant improvement (p > 0.05) for the more complex models over baseline U-Net. This suggests that architectural sophistication does not guarantee superior outcomes in breast region segmentation using thermal images.

A critical insight from this study is the effectiveness of simpler models. The baseline U-Net with ADAM optimizer consistently produced high Dice (0.9630), IoU (0.9292), and specificity (0.9801) while maintaining computational efficiency and stability. These findings highlight that in medical image analysis, especially with limited datasets, robust optimization and careful training can outweigh added architectural complexity. Thus, for clinical or resource-constrained applications, standard U-Net trained with ADAM offers the best balance between accuracy, interpretability, and computational cost, making it a practical and reliable choice.

This study makes a significant contribution to the field of thermography-based breast region segmentation by systematically evaluating and comparing the performance of three deep learning models—U-Net, U-Net with Spatial Attention, and U-Net++. The novelty of this research lies in its comprehensive analysis of the impact of different optimizers on model training, focusing on ADAM, NADAM, RMSPROP, SGDM, and ADADELTA. Beyond technical benchmarking, the study emphasizes dataset transparency, explicitly detailing the source, acquisition protocol, imaging device, and availability of the DMR-IR dataset, thereby ensuring reproducibility and reliability for future studies. A key finding is the superior performance of the baseline U-Net, particularly when trained with the ADAM optimizer. Despite being less complex than its variants, U-Net demonstrated high segmentation accuracy, interpretability through Grad-CAM, and reduced computational cost—highlighting that simplicity coupled with robust optimization can outperform architectural complexity.

From a clinical perspective, these results are highly relevant. U-Net’s ability to deliver strong precision and specificity reduces false positives, which is critical in breast cancer screening workflows. Meanwhile, the attention-based U-Net, with its improved sensitivity, may be suited to applications requiring the detection of subtle or ambiguous abnormalities. Together, these findings suggest that thermography, combined with deep learning segmentation, has potential as a low-cost adjunct to existing screening tools, particularly in resource-limited settings. This research contributes valuable insights into the selection of model architectures and optimizers for accurate and interpretable breast region segmentation in thermal images. The results provide a foundation for future research, guiding the development of advanced methodologies in medical imaging while also reinforcing the translational potential of thermography for clinical decision support. A key limitation of this study is that the manual annotations used to generate ground-truth masks were performed solely by the authors. Although cross-verification procedures were applied to minimize bias, the absence of certified radiologist annotations restricts the clinical validity of the segmentation masks. Future work will address this limitation by incorporating expert medical annotations to further strengthen reliability.