Deep Learning in Medical Image Processing – Scope and challenges

Medical Image Processing

Medical image processing is a set of procedures to obtain clinically useful information from various imaging modalities, mostly for enhancing diagnosis and prognosis according to the patient’s needs. It deals with the development of visual representations of raw medical image data for the purposes of further scientific analysis and medical investigation. In other words, medical image processing typically aims to extract features that might be difficult to assess with the naked eye.

Traditionally, the interpretations of medical data had been made by medical experts. However, due to the tremendous advancement in image acquisition devices, the data has become too large and complex for human experts to interpret and analyze. Besides, the elements of subjectivity, fatigue, noises, and variations due to various interference in data acquisition systems make it even more challenging for image analysis.

The solution is to use machine learning techniques to automate various medical image processing processes such as computer-aided diagnosis, image interpretation, image fusion, image detection and recognition, image registration, image segmentation, image retrieval, and analysis.

In recent years, machine learning (ML) and artificial intelligence (AI) techniques play a key role in healthcare, enhancing doctors’ abilities to diagnose and predict the risk of diseases accurately and quickly and prevent them in time.

These techniques consist of conventional algorithms like Support Vector Machine (SVM), Neural Network (NN), KNN, etc. and deep learning (DL) algorithms, namely Convolutional Neural Network (CNN), Recurrent neural network (RNN), Extreme Learning Model (ELM), Long Short term Memory (LSTM), Generative Adversarial Networks (GANs), etc. These algorithms try to automatically learn multiple levels of abstraction, representation, and information from a large set of images.

From machine learning to deep learning

Though automated detection of diseases based on conventional methods has been around for decades and new image acquisition devices have improved substantially over the recent few years, deep learning has emerged as the go-to methodology today to drastically enhance the performance of existing machine learning techniques solve previously intractable problems.

It has gained a central position in healthcare, especially in medical image processing, compared to traditional machine learning algorithms. Deep learning models halve the second-best error rate on image classification, and they have become the de facto standard for a wide variety of computer vision problems. These models are also not limited to image processing and analysis but can outperform other approaches in areas like natural language processing, speech recognition, and synthesis.

Deep learning is a part of machine learning and a special type of artificial neural network (ANN) that resembles the human cognition system. Deep learning approaches exhibit impressive performances in mimicking humans in medical imaging, especially in detecting structural abnormalities and classifying them into disease categories. They can overcome the limitations of previous computer-aided detection (CAD) systems such as false positives and can achieve greater detection accuracy, and help make human readers more productive by allowing them to shift humdrum, repetitive radiology tasks to AI.

Deep learning models can generate meaningful and powerful features after analyzing large amounts of uncategorized data and training the model for accurate prediction. They can adaptively and automatically learn a hierarchical representation of patterns, from low- to high-level features, and subsequently, identify the most important features for a given task.

Various types of deep learning algorithms are in use today in research. They include convolutional neural networks (CNN), deep neural network (DNN), deep belief network (DBN), deep autoencoder (dA), deep conventional extreme machine learning (DC-ELM), deep Boltzmann machine (DBM), recurrent neural network (RNN) and its variants like BLSTM and MDLATM, etc.


Modern deep learning methods have a high dependency on the quality and amount of training data. Therefore, the sheer amount of required training data is very high compared to conventional machine learning methods. Recent deep learning applications in brain MRI learned models from more than 1.2 million training data and the quality of the deep learning methods directly relies on the number and quality of training data samples. This is one of the biggest hurdles of deep learning research in medical imaging.

Many deep learning methods belong to the supervised approach. Therefore, they require manual labeling. Labeling thousands of imaging data is cumbersome. On top of that, inter and intraobserver variability also needs to be considered, making the labeling procedure even more problematic.

Another major issue related to sample size is the lack of balance in comparison groups such as the normal control group and the diseased group. Ideally, deep learning algorithms require an equal number of samples in the comparison groups. If there is a big imbalance between the comparison groups, the deep learning algorithm can not be able to fully learn the under-represented group.

Deep learning is a flexible modeling approach to learn an inherent representation of the input data. It can optimize a loss function to find millions of weights that can best explain the input training data. Therefore, it is very difficult to interpret how certain weights contribute to the final model. In some cases, we are left with a black box that performs the given task really well. In traditional machine learning, on the other hand, we can quantify how each semantic feature contributes to the final model and better understand or even improve the model as necessary. However, for deep learning, we cannot say the n-th weight in the j-th layer has a strong link with the final outcome. Such interpretability is largely lacking in deep learning.

Data interoperability and data privacy policies are other major barriers to be resolved before introducing deep learning methods in radiological practice. For instance, HIPAA provides legal rights to patients regarding their personally identifiable information and establishes obligations for healthcare providers to restrict or protect its use or disclosure.

Let’s sum up. Deep learning technology applied to medical imaging is the most disruptive technology that radiology has seen since the advent of digital imaging. Most researchers believe that within the next 15 years, deep learning-based applications will take over humans, and not only will most of the diagnosis be performed by intelligent machines but will also help predict disease, prescribe medicine, and guide in treatment.