The effect of hair removal and filtering on melanoma detection: a comparative deep learning study with AlexNet CNN

Related work

In recent research, significant progress has been observed in the use of CNNs and other artificial intelligence techniques for skin cancer classification. These approaches have shown great potential to help specialists reduce the time and resources required for diagnosis, which, in turn, allows for more appropriate treatment. A study by Leiter, Keim & Garbe (2020) compared expert opinion with artificial neural networks, and found that the software established a sensitivity of 95% and specificity of 88%, results comparable to those reported by dermatologists.

Several researchers have proposed specific approaches for skin cancer classification using CNN and other deep-learning techniques. Among these approaches, Yanchatuña et al. (2021) used a combination of CNNs and support vector machines (SVMs) to detect and classify skin cancer, obtaining an average accuracy between 80.67% and 90%, with an outstanding performance of 90.34% for the AlexNet plus SVM model. Popescu, El-Khatib & Ichim (2022) developed a skin lesion classification system involving multiple CNNs trained on the HAM10000 dataset, capable of predicting seven skin lesions, including melanoma. Ameri (2020) presented a deep CNN using transfer learning with AlexNet, eliminating the need for complex segmentation and feature extraction procedures by automatically learning useful features from raw images.

Other researchers explored the use of pre-trained networks (such as AlexNet and VGG16) in learning transfer and as feature extractors (Gulati & Bhogal, 2019). VGG16 with transfer learning was found to outperform other techniques in terms of accuracy and efficiency. This is attributed to its deep network structure and the ability to adapt to specific tasks through the use of pre-trained weights, resulting in better performance in the classification of skin lesions, as has been documented in various investigations (Anand et al., 2022). Alwakid et al. (2022) proposed an approach that incorporates ESRGAN, segmentation techniques, and a CNN along with a modified version of Resnet-50 for accurate classification. Similarly, hybrid models were presented that combine CNNs with SVM classifiers (Keerthana et al., 2023) or dimension reduction techniques such as PCA (Olayah et al., 2023). Naeem et al. (2022) developed the SCDNet model, which combines VGG16 with CNN and compares its accuracy with pre-trained classifiers in the medical domain.

In addition, comparisons of various CNN architectures were performed, where GoogleNet proved to be the most accurate (74.91% and 76.08%) in both training and test sets (Aljohani & Turki, 2022). Acosta et al. (2021) proposed a two-stage approach using CNN-based masks and regions and a ResNet152 structure to classify lesions as “benign” or “malignant”. Nida et al. (2019) introduced an automated melanoma region segmentation method based on a deep region-based convolutional neural network (RCNN) that accurately detects multiple affected regions. Gouda & Amudha (2020) presented LCNet, a model that requires no preprocessing or computation of specific features. Tahir et al. (2023) compared their novel DSCC_Net model with other deep benchmark networks, demonstrating improved accuracy.

Finally, Bansal, Garg & Soni (2022) introduced methods to remove hair in dermatoscopic images and used integrated features extracted using manual techniques and deep learning models to improve melanoma detection. Experimental results highlighted improvements in accuracy compared to approaches that do not apply preprocessing or use manual or deep learning features separately.

MSCD10000 dataset

MSCD10000 dataset stands for Melanoma Skin Cancer Dataset of 10,000 Images and consists of 10,605 dermoscopic images scaled to 300x300 pixels (Javid et al., 2023). These images were collected and resized from various publicly available ISIC (ISIC 2019, ISIC 2018) and HAM 10,000 datasets to create a balanced dataset with both benign and malignant classes. The dataset was downloaded from Kaggle at the following link: https://www.kaggle.com/datasets/hasnainjaved/melanoma-skin-cancer-dataset-of-10000-images. It is organized into two main categories, train and test. Within each category, the dataset is further divided into benign and malignant classes. Table 2 provides details on the distribution of dermoscopic images across these categories.

Hair removal and filtering

Before the hair removal process on dermoscopic images, we used four different filtering techniques to reduce noise and improve image quality: Fourier, Wavelet, Average blur, and Low-pass. Each filter was integrated using Python 3.x, cv2 libraries, and OpenCV, with our codebase adeptly adjusted to accommodate the unique requirements of each filter type. To illustrate, when implementing the low-pass filter, we applied a Gaussian blur with a 3×3 kernel size. PyWavelets for the implementation of Wavelet filters was used. The choice of these libraries allowed for efficient and reproducible processing of the images.

The process of filter implementation and hair removal is illustrated in Fig. 1. First, the images were converted to grayscale. They were then processed with a black top-hat filter and thresholding, followed by an inpainting method based on cv2.INPAINT_TELEA to restore the image (Telea, 2004). This process was performed using a 17x17 size core for the MORPH BLACKHAT operation, highlighting dark spots, which correspond mainly to hair, in the dermoscopic images (Ashraf et al., 2022).

After applying the filtering techniques, the grayscale images were subjected to a morphological operation for hair detection. We use a threshold value of 10 for binarization of the images, classifying them into pixels that represent hair (greater than 10) and background pixels (less than or equal to 10). The images were stored in separate folders labeled “with hair” and “without hair.”

The computational complexity of the hair removal process was considerable. On average, processing each image folder took between 2:30 and 3 h, with a memory requirement of approximately 6 GB. This time includes the application of the filters, the grayscale conversion, the morphological operation, the thresholding, and the filling method (inpainting). These factors were essential to ensure the viability of the method in both research and clinical practice.

Image pre-processing

The recommended optimal size for input images to ensure compatibility with the CNN architecture’s input layer is 224 × 224 × 3 (Krizhevsky, Sutskever & Hinton, 2012). The resized images are converted into a PyTorch tensor and then, the tensor values for each image are normalized sequentially with the mean and standard deviation values for each color channel. That is, scaling the values so that they have a mean of 0 and a standard deviation of 1 (LeCun et al., 1998; Glorot & Bengio, 2010; Ioffe & Szegedy, 2015).

Furthermore, data augmentation is used to address overfitting issues and improve the model’s capacity to generalize to new data by increasing the number of training samples for each class through random transformations not only for imbalanced datasets but also for datasets of any size (Perez et al., 2018; Ali et al., 2022). We applied different transformations, which included random horizontal flipping with a 50% probability, random rotations of up to 10 degrees, random adjustments in brightness, contrast, saturation, and hue within a maximum range of 0.2, as well as color channel normalization and conversion to the torch. Tensor format to effectively utilize the images as inputs.

AlexNet architecture

A visual representation of the AlexNet architecture used in this study is provided in Fig. 2. It was trained with hyperparameters carefully chosen to balance effective model convergence and computational efficiency. The learning rate was set to 0.001 to control parameter updates, while a momentum of 0.9 accelerated convergence by incorporating past gradients. A weight decay of 0.001 added regularization to prevent overfitting. The Cross-Entropy Loss function, suitable for multi-class or binary classification, was employed to measure prediction accuracy. A batch size of 64 determined the number of samples processed in each iteration, balancing memory constraints and training speed. The model underwent 30 epochs.

The initial layout encompasses five convolutional layers, each using specialized filters to identify specific features, such as contours and textures, present in the original image. These convolutional layers alternate with maximum aggregation layers, designed to condense the feature maps and simplify the recognition of more holistic patterns, particularly those that might remain unnoticed in larger images. The first, second, third, and fifth convolutional layers are seconded by maximum aggregation layers. An additional classifier module, consisting of three fully connected layers, is introduced to map the extracted features and thus generate the final binary classification.

In a variation on the original AlexNet structure, a new adaptive layer is incorporated between the convolutional and fully connected layers. This dynamically adjustable layer modulates the size of the output tensor to fit a standardized 6x6 dimension. This adjustment leads to an optimization of the model’s performance, allowing it to efficiently adapt to the diversity of features present in the medical images under analysis.

Performance evaluation

We employed the K-fold cross-validation technique for evaluating the average performance of the model in terms of the average training and test loss. The train directory of each dataset, original and modified, is divided into 10 folds of equal size. The choice of the value for k was determined based on the dataset’s size. The model undergoes 10 training iterations, each consisting of 30 epochs, with one fold serving as the validation set in each iteration and the remaining 9 folds used for training.

In addition, a quantitative evaluation of the CNN performance was carried out using four common metrics: accuracy, sensitivity, precision, and F1 score. True negatives (TN) and true positives (TP) refer to the accurate classification of negative and positive instances, respectively. On the other hand, false negatives (FN) and false positives (FP) refer to the incorrect classification of positive and negative instances, respectively.

${\displaystyle Accuracy=\frac{Truedetectedmelanomacases\left(TP+TN\right)}{Allcases}}$ ${\displaystyle Recall=\frac{Truedetectedmelanomacases\left(TP\right)}{Allmelanomacases\left(TP+FN\right)}}$ ${\displaystyle Precision=\frac{Truedetectedmelanomacases\left(TP\right)}{Trueandfalsedetectedmelanomacases\left(TP+FP\right)}}$ ${\displaystyle F1score=\frac{Precision\ast Recall}{Precision+Recall}.}$

Hair removal results

To ensure optimal model selection, we performed model evaluation every 30 epochs for each fold. Figure 3 shows the training and testing (validation) loss vs. epochs for the different datasets (original and modified) presenting a detailed analysis of the evolution of the loss throughout the training. The model curves are not smooth, instead, a lot of fluctuations are observed. These fluctuations could be attributed to several factors, with batch size being one of them. Notably, prior research suggests that a batch size of 32 often leads to optimal results in neural network training and as batch size increases, the convergence can be negatively impacted (Keskar et al., 2017). However, it is observed that the training loss is reduced to 0.1 for all of the model loss curves, which can be considered a good sign for correct segmentation results as other studies conclude (Ishida et al., 2021; Guefrechi et al., 2021).

The loss curve for the original images shows an increase in both training and testing loss at 30 epochs, which is not observed in the rest of the curves. Furthermore, even when data augmentation was used, some overfitting was observed when using the Fourier filter before hair removal, which can cause a decrease in generalization ability.

The overall CNN performance is determined by comparing the results of the four metrics between the original dataset and the modified datasets. The results in Table 3 reveal that the model performs the best among three of the four metrics with images where we have applied the Wavelet filter for noise reduction combined with hair removal. Here, the model achieves an accuracy of 91.30%, and a recall of 87%, while the low-pass filter yields a slightly higher recall of 89.80%. Additionally, when utilizing the Wavelet filter, the model has a precision of 95.19% and an F1 score of 90.91%.

The results indicate that hair removal improves the model’s accuracy and precision when it is combined with the Wavelet filter allowing the model to not only make correct predictions overall but also minimize false positives. Wavelet transform has already been used to equalize the noise created by fine hairs in dermoscopic images as well as to increase the algorithm’s detection speed in hybrid models combining deep learning and machine learning (Suiçmez et al., 2023). Some of the highest accuracy and sensitivity scores in the literature have been obtained when using Wavelet filters (Narasimhan & Elamaran, 2016).

On the other hand, the model trained on the original images exhibits the second-highest accuracy and precision of 91.10% and 94.19%, respectively. However, the second highest sensibility is reached when using the Fourier filter combined with hair removal.

These results are supported by the graphical representation of a box plot (Fig. 4), which visually points out the significant difference between the filters as a function of their performance. In Fig. 4, it is important to note that the Y-axis of the plot does not represent a specific metric, but rather the combination of four different metrics, thus providing a comprehensive view of each filter’s performance. The length of the boxes in the figure indicates the variability in the performance of each filter: those with longer boxes exhibit higher variability, while those with shorter boxes exhibit lower variability. This observation is supported by examining the confidence intervals, which are considerably wider for the Wavelet filter compared to the other filters. However, it is pertinent to note that, despite this superiority, the confidence intervals for the Wavelet filter are also the widest, indicating considerable variability in its performance. This observation raises the possibility that, although the Wavelet filter excels in general terms, its performance may fluctuate significantly in different scenarios, which is why we consider that further investigation into the robustness and generalizability of the filter’s performance in diverse settings is needed.

Furthermore, it is important to note that the significance of this improvement in relation to variations in the initialization of network parameters or other potential factors is not extensively discussed since hyper-tuning techniques were not applied to the network used in this study. It could be suggested to continually evaluate and potentially fine-tune these hyperparameters and initialization methods based on the characteristics of the melanoma dataset to ensure robust training and accurate classification results.

Overall, these outcomes strongly suggest the model’s competence in effectively distinguishing between benign and malignant skin lesions, as evidenced by its robust performance metrics. CNNs offer significant advantages for physicians in clinical settings since they can process large volumes of data quickly and consistently, continuously learn new information, and offer objective assessments (Krishnan et al., 2023). For instance, the study by Haenssle et al. (2018) compared a CNN’s diagnostic performance with that of a large international group of 58 dermatologists, revealing the CNN’s superior performance. The CNN exhibited a higher specificity (82.5%) compared to dermatologists at both level-I (71.3%, P < 0.01) and level-II (75.7%, P < 0.01) sensitivities of 86.6% and 88.9%, respectively. Nonetheless, experts maintain an edge in handling complex or ambiguous skin lesion cases.

Comparison with related work

We validated our model’s performance against two setups: Wavelet filter with hair removal (Proposed 1) and original images without hair removal (Proposed 2). These results are contrasted with some well-established recent studies in the field shown in Table 1. As can be seen in Table 4, the proposed model was optimal in terms of the four metrics: accuracy (AC), recall (SE), precision (PR), and F1 score (Fsc). It has one of the highest accuracy and precision values.

Single-factor analysis of variance (ANOVA) was applied to investigate the existence of statistically significant differences in performance among various neural network models. This method will provide a comprehensive perspective on whether at least one model is statistically different from the others with respect to the performance metrics under evaluation. The results, obtained through Statistix 10 software and presented in the table below, reveal an F-statistic of 4.55 with an associated p-value of 0.0042 as shown in Table 5.

The associated p-value being below the critical threshold of 0.05, leads to the rejection of the null hypothesis. Consequently, the existence of at least one significant difference in performance between the various neural network models is established.

Overall, the proposed model exhibits competitive performance when compared to the referenced studies. It achieves similar levels of accuracy, recall, precision, and F1 scores, emphasizing its efficacy in detecting melanoma in dermoscopic images. However, it is important to acknowledge that there is still room for improvement to develop and compare new filters for noise removal that might enhance hair removal outcomes in the field of melanoma detection.