Machine Learning in Healthcare: Cancer Detection

Skin cancer is a prevalent and wide-ranging health concern, affecting millions of individuals globally. As with most illnesses, early detection of cancer significantly improves the likelihood of successful treatment. Because of this, it is important to diagnose as quickly and accurately as possible. Machine Learning in healthcare has started to play a pivotal role in this regard.

Dermatologists traditionally rely on visual inspection for diagnosing skin conditions. However, the human eye has limitations, especially when dealing with subtle or complex patterns. The various categories of skin lesions can look very similar to a person, and especially under time pressure a misdiagnosis can be easily made. 

Machine Learning in Healthcare:

Skin Cancer Detection

Overview of Image Classification 

Image classification is a machine-learning technique that assigns images to a set of predefined labels or categories. This powerful tool has found applications across various fields, including medicine, where it proves invaluable in automating and augmenting diagnostic processes. 

With the current advances in machine learning and especially neural networks, image classification is more suited than ever for faster and potentially more accurate classification of potential skin cancer cases. By using machine learning in healthcare, we can negate many disadvantages of traditional early cancer screening methods and invite new advantages, saving lives.

One of the primary advantages of employing machine learning in skin cancer diagnosis is the potential to save time and reduce the strain on healthcare resources. Using machine learning in healthcare allows for Automated image classification would be nearly instantaneous and even if just serving as a first opinion it would free up time enabling healthcare professionals to focus on more complex cases and patient care. 

Finding a Dataset 

The foundation of any machine learning project is to find a dataset with good quality and quantity. You need enough good-quality data that the machine learning algorithm can learn the differences between the categories it needs to distinguish. For this project, the HAM MNIST 10000 dataset is available from Kaggle, which is particularly beneficial for developing a Skin cancer detection AI. This dataset includes over 10,000 images of skin lesions with their associated diagnosis, providing a diverse and comprehensive set for training and evaluation.   

The dataset includes examples of seven of the most common categories of skin lesions: 

• Actinic Keratoses: Precancerous growths induced by sun exposure. 
• Basal Cell Carcinoma: A common form of skin cancer found in sun-exposed areas. 
• Benign Keratosis-like Lesions: Non-cancerous skin lesions, including benign warty growths. 
• Dermatofibroma: Benign, hard nodules, often on the lower legs. 
• Melanoma: A malignant form of skin cancer originating from melanocytes, which is one of the primary focuses of machine learning in healthcare.
• Melanocytic Nevi: Common and benign moles or nevi. 

• Vascular Lesions: Non-cancerous blood vessel tumours and port-wine stains.

As well as including the images, the dataset has additional patient information such as age, gender, and the specific location of the lesion. This additional data can also be used by the image classification model, potentially increasing its accuracy. For machine learning in healthcare, such comprehensive data is invaluable. 

Exploring the Dataset 

As with any data-based project, the first step was to perform exploratory data analysis (EDA). This foundational process is crucial for machine learning in healthcare, you go through the dataset looking for quality issues and interesting high-level insights. During this, I noticed that the dataset is predominantly composed of melanocytic nevi (moles). Given the widespread occurrence of benign moles, this composition aligns with expectations and mirrors patterns often encountered in skin cancer detection AI projects.

However, the prevalence of one category raises concerns about dataset imbalance. Addressing this imbalance is critical to prevent biased model training, as without it the model may just preferentially predict the majority class as it finds this gives good accuracy. As an extreme example, imagine a dataset where 90% of the images were of one category, by always predicting this class the model would achieve 90% accuracy. This model has not learnt what we wanted and is very biased. The main ways of addressing this are to resample the dataset to balance it, assign different weights to each class, or use a different performance metric. 

Another noteworthy aspect is the dataset’s predominant representation of images of Caucasian skin. This limitation prompts consideration of potential challenges in accurately classifying skin lesions in individuals with darker skin tones. 

Convolutional Neural Networks 

Convolutional Neural Networks (CNNs) work especially well with image data, making them the architecture of choice for image classification tasks. As the name suggests, CNNs employ convolutional layers, which extract patterns from images. 

Pretrained Models  

With the dataset being very large and made up of images, which hold a lot of data, the training process for this task can be quite lengthy. To speed it up, pre-trained models such as VGG16 and MobileNet can be used. These have been trained for general image classification on much larger datasets. Although they haven’t been trained on the specific topic of image classification, much of their training is still useful. Transfer learning can be used to allow these models to be slightly adapted for the task of skin cancer classification. By transferring their general feature extraction capabilities whilst training them to distinguish the categories for this task, they proved a very useful tool for quickly building effective models. 

Custom Model Architectures 

While pre-trained models offer training efficiency, I still wanted to try some custom architectures that I could tailor to the specifics of this task. Because the dataset has additional information about the lesion, such as patient age, sex, and lesion location, I wanted to utilize multi-input neural networks. These combine a convolutional and regular deep neural network, allowing all of the information in the dataset to be used. 

Grad CAM: A Visual Insight into Model Interpretation 

When using machine learning in healthcare, model transparency is a very common issue. It is often unclear as to what pattern the network is using to classify the input. For these images,  it is important to check that the model is focussed on the section of the lesion with the images, as this gives an indication that its diagnosis is based on the lesion. A critical aspect of deploying these models is ensuring that they focus on the lesions in images. Grad CAM (Gradient-weighted Class Activation Mapping) serves as a valuable tool to assess where the model is directing its attention within the image. 

Making the Models Accessible and User-Friendly 

To make these powerful models accessible to a broader audience, I created a Flask web application to host it. Flask serves as a framework to showcase the various neural network models and allows users to assess the performance of each model. 

The user interface then allows them to upload their images of skin lesions, select their preferred model, and receive a classification of the uploaded images. This approach simplifies the user experience, making it much more accessible than running models through the Python scripts I used. 

Machine Learning in Healthcare: Future Work 

Recognizing the importance of diverse representation in medical datasets, future work for machine learning in healthcare involves finding additional data that has a broader range of skin tones. This would improve the model’s generalization across different demographic groups, making it a useful tool for everyone. 

The models in skin cancer detection AI are currently achieving accuracies of around 70%, so continual improvement remains a priority. At this level of performance, they are not yet reliable enough for practical use, underscoring the necessity for ongoing refinements for machine learning in healthcare through model architecture adjustments and targeted training. However, this does serve as a valuable proof of concept, and I firmly believe these techniques will be used in the future. 

The use of machine learning in healthcare for image classification is a very promising tool that could save healthcare professionals a lot of time. Using a large, publicly available dataset I was able to adapt pre-trained models and design custom architectures to classify various types of skin lesions. This was then made accessible through a user-friendly web interface, providing a very promising proof of concept that this technology will be widely available in the future. There is still some way to go and the development of these models is of paramount importance as their early classification has the potential to save lives. 

Interested in joining our diverse team? Find out more about the Rockborne graduate programme here.