Skin Cancer Classification with Deep Learning

Key Takeaways

[Music] this video will present dermatologist level classification of skin cancer with deep neural networks convolutional neural networks are a branch of deep learning that Maps images to class labels of really popular way of understanding this is through examples of classifying images of cats and dogs as shown in this image the convolutional neural network takes as input an image and then uses a series of convolutional filters and other miscellaneous things that the deep learning community has come up with for computer vision applications and then Maps the images into class labels such as cat or dog the idea here is to use the same kind of idea but to detect skin cancer in in clinical images so there are 5.4 million new cases of skin cancer every year that account for 10,000 annual deaths but early detection of this is critical and that's why this idea of being able to embed a dermatologist level classifier into your phone so you could just snap a picture of your skin lesion and instantly have this kind of diagnosis is a very appealing sort of idea so the data set that they use in the paper consists of one hundred and twenty nine thousand four hundred fifty clinical images including 3374 Mosca p images and this data said showed on the lobby's this images showed them the left of the slide is sort of what this data set would look like so what they use to classify this is they start with transfer learning and transfer learning is a really popular technique in deep learning that's why I've pictured the Swiss Army knife at the bottom of the slide transfer learning is frequently used when you have limited datasets so in this case a hundred thousand images may seem like a lot but it's really not that much for these deep neural networks which are very data data intensive so the idea of transfer learning is you take a pre-trade network such as the Google Met inception v3 and you train on a big data set like image net which is one point two eight million images in the thousand object categories then the final layer the classification layer is removed and replaced with a new layer for the new tasks so one other interesting issue of using transfer learning is that the skin lesion images have to be resized to match the original input of the original dataset so imagenet images of resizes 299 by 299 and when you so now when you use the skin lesion images you have to resize them to that same input dimension so the novel insight in this research paper is to construct a taxonomy of the output structure and this is another idea that's been seen in papers like word Tyvek with these hierarchical output spaces so the idea is that the convolutional network doesn't just directly predict the classes rather it predicts probabilities for each of the classes and then these classes are traversed upwards and summed up to make the predictions so you would use all the leaf nodes under leaf nodes under something like benign to aggregate the prediction to the final prediction so you sum up the little leaf nodes here things like lipoma fib roma and cysts you sum up all the leaf nodes and then aggregate that to be the benign fiction and each of these high-level classes like the line malignant and non neoplastic have their own sets of leaf nodes and also interestingly they have different cardinalities of leaf nodes like in this example the non neoplastic higher level node has many more leaf nodes than benign and malignant so this is the high-level overview of what it looks like a skin lesion input is images inputted into this Google net inception Network which has this structure of branching the previous input into these four separate convolutional blocks and then aggregating them together and then it is mapped into this taxonomy output space to make the predictions so this is a t-sne visualization of CNN features what a t-sne visualisation does is it lets you go from the very high dimensional and completely impossible to visualize 2048 dimensional feature layer before the classifier classifier output this prediction and you can project this into a low dimensional representation so we can kind of see what classification decision boundary the classifier might be able to do with the deep learning feature extractor so in this case you can see that there is a pretty clear separation of the orange blue green and yellow classes but there still is a significant overlap you can see a lot of blue in the red and yellow and the green so these are the results from the CNN compared to the dermatologist it outperforms the dermatologist by a good margin on the three-way classification tasks where you are predicting the high level notes by traversing up from the leaf nodes but then when you predict in the leaf nodes it performs the same as the dermatologist this is the confusion matrices of the nine way classification problem and it's sort of interesting because the dermatologist and the CNN they have some similar biases in their missed predictions you see that predicting six when it's really zero is a common trend amongst the CNN dermatologists one and dermatologists too when you deal with class imbalance datasets like in this case I don't remember the exact dimensions but there might be like 130 melanoma images 270 benign or you know some distribution that is skewed towards one class so the way that you would account for this is by using these metrics that account for imbalance so sensitivity is where you take the true positive rate and divided by the positives specificity is the true negative by the negative and this AUC area under the curve shows the trade-off between the two as you change the decision boundary so you might say if it's you know if the prediction lies above a certain value like 0.5 then you predict it one way or the other and you can slide that up and down to optimize for sensitivity or supposed specificity so these are the saliency maps which is another common practice in computer vision to try and highlight what the CNN is looking at to make its prediction and it's pretty encouraging because it is accurately looking at the at the skin lesion rather than arbitrary parts of the image thanks for watching this video on skin cancer classification with deep learning you can get the full paper link in description after watching this video I highly recommend checking out this explaining in detail the inception that architecture that they use thanks for watching Henry AI Labs please subscribe