Effective Diagnosis of Melanoma Using Artificial Intelligence

Introduction

Artificial intelligence is the ability of a computer or computer-controlled robot to perform tasks commonly associated with intelligent beings. When asked about AI the average person may immediately imagine the Hollywood depiction of AI such as a Terminator, however, this is a grand exaggeration of current AI ability. AI begins with a program that learns how to perform basic tasks it has been instructed with and with each succession of its task, the task will become more difficult. This “trains” the AI on how to calculate or complete more complex tasks over time. Depending on what task the AI has been challenged to complete, the resulting AI may be more specialized in a particular task. Although there have been large strides of progress in developing computer processing speed and memory capacity, there is no AI programs that can match a human’s flexibility over wider domains or in tasks that require everyday knowledge. In specializing in one specific category, however, AI can match the performance levels of human experts such as medical diagnosis, search engines or voice and handwriting recognition. The use of AI has made its way into medical science areas and has since been used in a range of different medical fields to promote therapeutic development. AI has since been used extensively in a range of medical fields such as diagnosing acute and chronic diseases. In this critical essay, we will focus on the use of AI in Melanoma detection, with references to Cui, et al’s research in assessing the effectiveness of artificial intelligence methods for melanoma. In this research, their objective is to determine the most effective AI method for diagnosing melanoma. Overall, the use of AI would be a significant benefit in the use of diagnosing diseases such as melanoma, given the accuracy of its results, however, this does not mean it should be used blindly to diagnose every patient, rather human experts' supervision should still be an integral part in the diagnosis.

Summary

The use of AI in the medical science field has grown extensively to include the diagnosis of diseases. Research in assessing the most effective AI method for diagnosing melanoma will be the focal area in which this review will revolve around.

The method seemed adequate, where both AI training methods were tested, traditional machine learning and deep learning algorithms, against the same sample data sets, one large set consisting of 2200 images of which 564 were melanomas and the rest were non-melanomas, and one subset containing 606 images of which 295 were melanomas and 311 were nonmelanomas, ensuring reliable and valid results.

In assessing the traditional machine learning AI, the use of 4 fully automatic image segmentation algorithms was used appropriately for segmenting melanoma. However further explanation as to what dilating, corrosion, and hole filling could be provided to further explain image processing processes. The morphological features, area, perimeter, roundness, centroid, and many others were calculated as texture features. From this post-processing, the least amount of shrinkage and selection operator method used to select the most useful features was a good decision as it ensures relevant data is being used. A RAD score calculated for each image is a good measure of a linear combination of selected features (SVM, classification and regression tree, k-nearest neighbors, and LR) weighted against their respective coefficients.

From these values, the accuracy, specificity, and sensitivity were calculated using a 10-fold cross-validation method. This method was appropriate as it estimates the skill of the machine learning models. The region selected as important by dermatologists were compared to the regional segments selected by the four different algorithms to determine the intersection over union, which is a statistic used for gauging the similarity and diversity of different samples, and the false-positive rate.

Of the four image segmentation algorithms, the best as shown above was Region growing. This algorithm was then used as the base algorithm for traditional machine learning algorithms. The algorithm chosen was an acceptable decision as it provides the best results of each AI.

The outcome of the best AI using the choice algorithm in traditional machine learning was determined to be the LR algorithm. The LR algorithm produced the highest accuracy, comparable to SVM, but it has a higher sensitivity rating at the sacrifice of some specificity, thus making it a better algorithm for an all-rounder type.

Deep learning methods process data in an entirely distinct method, where deep learning requires a large amount of image data, skin cancer images were fewer than needed. To compensate for the lack of data, transfer learning was used to solve the problem, this method essentially uses an experienced or trained model which then transfers relevant data to the newer models, reducing the overall amount of time and resources to train the newer models. This method is questionable as it gives the deep learning AI programs an edge in diagnosing melanoma, as it has had a larger data set to gather information and process compared to the traditional machine learning. Nonetheless, this does give the Deep learning AI a chance to use its full potential and provide the best data values. The four most commonly used deep learning AI were challenged to diagnose the same images as the ones used for traditional machine learning, and from Table 3 Google’s Inception V3 AI was able to produce the best results, with the highest average accuracy, sensitivity, and specificity.)

These results in Table 3 provide promising benefits for the medical field, as the depth of knowledge acquired by the deep learning AI and the increase in training time so too does the accuracy and classification of the experiment. Although the experiment appears sound overall, there are limitations throughout the experiment. One of which was the severity of the melanoma was not determined, another was the nonmelanoma samples were all of the benign nevus otherwise known as birthmarks, this particular limitation is of great significance as the results may not necessarily mean the AI was able to diagnose melanoma, but rather the absence of melanoma. To accommodate this flaw, the nonmelanoma images should contain a variety of skin conditions, to ensure the AI is specifically diagnosing melanoma.

Conclusion

There are advantages and disadvantages that ultimately determine the appropriate use of AI. Traditional machine learning algorithms, they require relatively low data sets and is easy to train, however, this comes at the cost of low generalization ability and high costs. Deep learning AI is able to be more flexible and include more data sets and contains extraction features although these require much larger data requirements and higher quality data. Although deep learning AI has proven to be accurate, it should still be used under the supervision of expert dermatologists to confirm diagnoses, because the data used to “teach” the AI was limited in certain areas. Overall, the assessment of effective diagnosis of melanoma using artificial intelligence was well done, nevertheless, there are improvements that can be made to improve results.