The client is a reputed diagnostic center spread across 10 locations in the US.
Problem Statement & Challenges
A considerable quantity of X-ray images was being captured daily by the client and had to pass through manual methods of medical analysis by a radiologist who interpreted each of these images to diagnose the diseased condition of patients. This was a time-consuming manual process and happened to be the primary means of medical analysis which posed a significant challenge for both clinicians and patients. Due to shortage of radiologists, analyzing the vast quantity of images proved to be a challenge, further increasing the turnaround time for patients to get their X-ray reports. On average, it took 50 minutes to generate a report, and it took approximately 7 hours for the final report.
- Decentralized data or local repositories
- Increased turnaround time (TAT) due to manual methods of analysis
To build a sophisticated ‘Convolution Neural Network’ to automate the X-ray analysis by classifying the images into categories comprising of reasonable and abnormal conditions. Before handing over the selected cases to the radiologist for a detailed interpretation, this solution should
- Be real-time
- Should act as an initial screening
Approximately 1,30,000 X-ray images were collected to train the Machine Learning model. The X-ray images collected were passed through an image process for enhancement of the diagnostic information and suppressing of the non-diagnostic information. This input image data was driven through multiple sets of convolution layer and pooling layer before the flattened data was distributed to the ANN classifier. We utilized a high-level API called Keras with Tensorflow backend to build a neural network which served as a solution.
- Developing a process to collect the X-ray images and loading them to a centralized server system from disparate sources
- A real-time execution analytics engine which executes the image analytics comprising of Machine Learning algorithms that consumes these X-ray images, processes the images, classifies them into different categories.
Results / Impact
- Centralized data management system for all the images from various centers
- Digitization and automation has cut costs and eliminated errors which resulted in significant dollar savings
- Standardization of the medical image analysis process across all the centers of the client
- The image analytics model achieved an accuracy of 92% (the model is frequently retrained online at client’s location with new data to ensure enhanced accuracy)
- The turnaround time (TAT) decreased by 54%
Considering the shortage of radiologists, an automated solution in medical imaging with a reasonable degree of accuracy has proven to improve their productivity. This solution helps patients receive early diagnosis and treatment, decreasing mortality rates.