Fig 1: Segmentation on SISS dataset.
The purpose of this project is to build a CNN model for stroke lesion segmentaion using ISLES 2015 dataset.
Recent studies have shown the potential of using magnetic resonance imaging (MRI) in diagnosing ischemic stroke. Reviewing hundreds of slices produced by MRI, however, takes a lot of time and can lead to numerous human errors. It is widely accepted by the medical practitioners that automated segmentation methods for ischemic stroke lesions could significantly speed up the beginning of a patient’s treatment. The automated segmentation can locate the tissue with lesions and give an estimate of its volume, which helps in the clinical practice by providing a better assessement and evaluation of the risks of each treatment. These reasons highlight the need for a fully automatic ischemic stroke lesion segmentation approach using a flexible, fast and effective deep neural network.
File: There are 4 types of MRI scan for one person
File Format: .nii
Image Shape: 240(Slide Width) × 240(Slide Height) × 155(Number of Slide) × 4(Multi-mode)
Image Mode: 4 (Multi-mode)
Fig 2: SISS dataset.
File: There are 7 types of MRI scan for one person
File Format: .nii
Image Shape: 96(Slide Width) × 110(Slide Height) × 71(Number of Slide) × 7(Multi-mode)
Image Mode: 7 (Multi-mode)
Fig 3: SPES dataset.
| Methods | Range |
|---|---|
| rescale | 1.0/255 |
| rotation_range | 30 |
| horizontal_flip | True |
| vertical_flip | True |
| shear_range | 0.2 |
| zoom_range | 0.1 |
Fig 4: The architecture of the proposed framework for Ischemic Stroke Lesion segmentation.
Fig 5: An schematic diagram of the proposed Inception block.
True Positive (TP): TP implies number of true positives, that is, positive correctly identified as positive.
True Negative (TN): TN implies number of true negatives, that is, negative correctly identified as negative.
False Positive (FP): FP implies number of false positives, that is, negative incorrectly identified as positive.
False Negative (FN): FN implies number of false negatives, that is, positive incorrectly identified as negative.
Lri = Initial Learning Rate = 0.0001
decay factor(f) = 0.2
step = 2
Fig 6: Performance of proposed network in term of dice coefficient on each modality on SISS and SPES Dataset for various loss functions.
Fig 7: From left to right: Plot for loss DSC and accuracy for training and validation set on SISS dataset.
Fig 8: From left to right: Plot for loss DSC and accuracy for training and validation set on SPES dataset.
This study models a modification of the U-Net and Inception Net trained upon a well-known public dataset ISLES 2015, proposed for the automatic segmentation of stroke lesion from MR images. Upon careful evaluation on the above dataset, our model demonstrated an increase in the efficiency and performance, outperforming some of the other modern CNN based architectures. The same can be attributed to the flexibility of selecting the training method and the structure of the proposed model which in turn enhanced the training stability. The model has been tested for various loss functions for different modalities. A patient-wise study led to the conclusion that there is a further scope for improvement while detecting very small lesions. For subsequent work, we aim to expand onto other lesions studying different disorders as we believe that our model holds great potential in the medical field in relation to the diagnosis and treatment of several disorders.