Artificial Intelligence (AI) in Early Detection of Skin Cancer

AIM

Modern artificial intelligence is used for the diagnosis of skin cancer and for the classification of skin cancer; at present, the techniques and methods of AI are applied in the detection of skin cancer. The accuracy, specificity, precision and sensitivity of AI tools. The advantages and limitations of using AI for the diagnosis of skin cancer.

METHODOLOGY

1) The use of deep learning algorithms for identifying and segmenting skin lesions. One of the most important aspects of applying an AI system in the detection of skin cancer is to carry out the detection and segmentation of skin lesions. In order to achieve accurate detection, all potential cancerous growths must be detected and identified in the image, while segmentation refers to the precise determination of the boundaries of the lesion. This is particularly important for the extraction of particular features that can be used for classification. Accuracy is essential for both the detection and segmentation of lesions, and a number of artificial intelligence algorithms are employed in the detection of skin lesions.

2) Deep learning algorithms which are based on convolutional neural networks have thus been shown to be effective for efficient object detection in medical images. In order to identify the initial area of interest, the K-mean clustering part of these algorithms is combined with the extraction of relevant features and the use of ensemble learning techniques in order to detect the presence of lesions. The system is able to classify a specific pitted hole lesion according to the pixel intensity value. These deep learning methods are expected to work best when it comes to automating the learning of complex features directly from image data, especially in the case of lesion-detection tasks.

3) U-nets are widely accepted and well-justified deep learning models when it comes to skin lesion segmentation in the field of medical imaging. The U-net structure is made up of down-sampling and up-sampling paths.

4) The path used for down-sampling encodes the input image into a lower-dimensional representation.

5) The up-sampling path carries out the reconstruction of the segmented images.

6) In order to train a U-net model for skin lesion segmentation, it is necessary to train the model using skin lesion images together with the corresponding ground truth segmentation masks so that the model will be able to tell the lesion area apart from the normal skin and the background.

There are some more tools that identify skin lesions, these are:

 a) Dermo-doctor

 b) DermoExper

 c) DSNet

7) Reinforcement learning

It is an alternative and innovative method of segmenting skin lesions, consisting of the intelligent agent learning how to segment the lesion area by means of a Markov decision process (MDP), with the agent receiving rewards for each of its actions. Using deep reinforcement learning algorithms such as the deep deterministic policy gradient (DDPG), the intelligence is gradually trained to enhance the segmentation results, progressing from a rough initial segmentation to more detailed boundaries.

Application of convolutional neural network (CNNs) for skin cancer classification

a. It has already been proven that Convolutional Neural Networks will most likely result in the best and most dominant type of architecture for use in medical image analysis; in particular, their most important application is the classification of skin cancer. This kind of architecture is able to learn complex hierarchical features directly from the image data and is therefore very well suited to detecting patterns that indicate malignancy in dermatopathological images.

b. In fact, CNNs are extremely accurate tools having been shown to classify skin cancer with a high degree of precision and have, on various occasions, been reported to outperform those of boar-certified dermatologists. They thus seem to be suitable as tools for self-diagnosis.

c. A multiple-specific CNN structure has been successfully applied to the task of classifying skin cancer. Among the more well-known architectures are Google Net, Inception V3, VGG-16, UGG-19, and Mobile Net. Each of these has its own particular strength and feature. Numerous studies have shown that it is possible to obtain very promising results with models such as Dense Net 201, which achieves an accuracy of 97% on the HAM10000 dataset, or with Inception U3, which attains 90% accuracy on that same dataset.

The researchers have looked at some other advanced hybrid CNN models in order to obtain better performance. In fact, they have managed to reach a very high average accuracy of 97 per cent in the classification of melanoma, basal cell carcinoma and squamous cell carcinoma. An interesting point is that the combination of LBP and CNNs has resulted in the highest value for hybridization in terms of accurate classification. Indeed, such accurate classification will be essential for formulating appropriate treatment strategies and for determining patient outcomes according to the types of skin cancer diagnosed. Nevertheless, the researchers have examined some modified and hybrid CNN architectures with a view to assessing their performance. The hybridizations of CNNs with Local Binary Patterns are inherently used for achieving high accuracy in classification and has been able to attain a very high average accuracy of 97.29 per cent in the research carried out on the classification of melanoma and basal cell carcinoma as well as squamous cell carcinoma. The ability of CNNs to distinguish between these different types of skin cancer is important since it helps in guiding the right treatment and in improving clinical outcomes.

While CNNs have produced promising results when it comes to classifying skin cancer from images, there has been a growing amount of research into other deep learning architectures such as RNNs and Transformers. Specifically, the modified version of RNNs, namely Long Short-Term Memory (LSTM) networks, are suitable for use with sequential input. LSTM networks have been studied in the context of skin cancer diagnosis by considering the temporal evolution of a skin lesion over time, even though this is not the typical application of RNNs but rather the use of CNNs for image classification. Similarly, hybrid systems that combine a CNN with other network architectures (CNN-RNNs) have been developed, these systems benefiting from the strengths of both architectures by having the CNN extract the spatial features of images and the RNN model the temporal dependencies. Some studies have reported very high accuracy being achieved using CNN-RNN models for multi-class skin cancer classification. Since classical RNNs are not very good at carrying out this kind of task, they encounter a number of problems due to their inability to capture the long-range dependencies that exist within images and also because of the vanishing gradient problem.

It will eventually happen that the New Transformation architectures, particularly Vision Transformers (ViTs), will become the foremost contenders replacing CNNs in the field of medical image analysis, including the classification of skin cancer. The true reason for the success of these achievements is natural language processing, since Transformers show how various parts of an input sequence influence the model's predictions through self-attention by assigning different levels of importance to them. Moreover, ViTs are able to model long-range dependencies in an image, a feature that could be valuable when analyzing complex skin lesions since different areas may be subtly connected. The properties of Transformer architectures also allow the entire input set to be processed in parallel, leading to Chetan Faster energy efficient training. Numerous studies have so far achieved very promising results regarding transfer learned multiclass skin cancer classification performance using pre-trained ViTs, with some even outperforming traditional transfer learning models based on CNNs. Because ViTs are robust to different types of perturbations in the input data, they also offer a high degree of architectural flexibility. The Transformer-based architecture is not only being explored for classification but also for other interesting applications in the diagnosis of skin cancer, such as lesion segmentation, lesion detection, and the extraction of semantic information from dermatoscopic images. The bright future prospects that the emerging transformers hold for these areas will mainly consist of opening up new promising avenues in skin cancer diagnosis that could be AI-driven and be beneficial globally in terms of context and precision.

d. Performance Metrics and Evaluation of AI Techniques Applied for Skin Cancer Detection

To measure performance of AI technology on skin cancer diagnosis-old evaluations rely on that it is knowable that there are number of performance measures each of which measures different aspect of machine prediction ability and classification of skin lesions. Some performance measures are (but not limited to) Accuracy which gives the fraction of all predictions made correctly and also the recall or sensitivity which measures machine ability of correctly identifying positive instances (malignant lesion) and specificity which measures machine ability to diagnose negative cases (benign lesion). Others measure that have surfaced on many occasions also include precision which gives the number of true positives out of all who are predicted positive and F1 score which is harmonic mean of precision and recall and gives a measure of accuracy balanced between the two parameters.1 AUC-ROC measures the overall performance of a classifier to accurately classify the outputs over any particular discrimination threshold. At times AUC-PPV-Positive Predictive Value and AUC-NPV-Negative Predictive Value has surfaced in consideration. The performance measures give a more overall measure of the system regarding the diagnosis on many other parameters while comparing the same on malignant or benign lesions.
In many studies, the performance of the AI model and dermatologist had been compared. A number of AI systems have been proven to perform better or equivalent than skin specialists in diagnosis of melanoma. AI algorithms reported high ROC values (>80%) in diagnosis of melanoma with dermoscopy images, while in general case sensitivity and specificity reported were 83.01% and 85.58%. In several situations, AI CAD systems are better than board certified dermatologist on accuracy in detection of skin tumours, even with small data. It is known that accuracy of human skin specialists for detection of melanoma on average is ranging from 60%-70% and it can easily out beat with the models when measured under lab conditions. The performance may vary from different types of reasons such as type of AI method used, quality and size of training data-sets, actual diagnostic task difficulty. A number of published studies fail to use publicly available data set for training and testing, and evaluation methods are quite different, which add difficulty for better comparison between methods. In any case, the AI models are becoming to demonstrate very good performance in distinguishing different skin tumours, and it relies on many CNN models and Vision Transformer methods for identifying with higher accuracy regarding melanoma. Other commonly occurring tumours such as basal cell carcinoma and squamous cell carcinoma showed promising results for different performance measures.