Session Item

The aim is to demonstrate the ability of a neural network based on the TensorFlow software to perform automatic delineation of the hippocampus based only on lateral and frontal projections of the brain contour obtained from a CT-scan.

The computer used was a 64 bits Intel Core i5 desktop computer with Python 3.8.3 and TensorFlow version 2.2.0 installed. The neural structure used consisted of three stages. First three dense layers with a total of 3940 neurons. Then a reshaping of the data, and finally a transposed convolutional layer that generated the output image. The data basis was the delineated brain from CT-scans and the left hippocampus from MR-scans of 16 brain patients. Several Python programs were designed to do the following three types of data processing.

Training data

From the data basis the training data was created as a multitude of image triplets consisting of two input images and one output image. These triplets were made by rotating and longitudinally translating the 3D contours of the brain and left hippocampus in a multitude of ways. And register frontal (x-z) and lateral (y-z) projections of the brain and the corresponding transversal (x-y, z=0) contours of the left hippocampus. The brain projections were pooled to 100 × 100 pixels, smoothed and contrast enhanced. The hippocampus contour was filled it had a size of 56 × 56 pixels. (see figure 1, top row). The final number of training image triplets were 83871.

Training

The neural network was trained (fitted) to the relation between the input brain projections and the output hippocampus contour. The network was fitted for 2000 iterations which took 80 seconds each, close to two days in total.

Prediction

The trained neural network can predict a 2D hippocampus contour from two input projections (see figure 1, bottom). For an unknown patient several predictions can be joined to create a 3D hippocampus contour. This contour can be written in DICOM to be exported to the Eclipse treatment planning system.

Five brain patients unknown to the neural network were used for evaluation. The predictions of the left hippocampus were imported to the Eclipse treatment planning system (figure 2). The dice coefficients between the humanly delineated hippocampus (purple) and the machine delineated hippocampus (yellow) were found to be in the range from 0.15 to 0.41. This could be clinically relevant where only an approximate location of the hippocampus is needed.

It is found that a simple neural network based on the TensorFlow software package is able to predict the position and shape of the hippocampus fairly well based only on the brain contour from a CT-scan. This find is useful in itself because a MR-scan can be omitted, but also promising for future developments of more sophisticated neural networks for automatic delineation.