• No results found

This thesis presents four strategies to discover and predict the various regions of an ischemic stroke using a dataset of CTP images; however, it could be possible to improve the achieved results with future researches. An interesting study for the future could be to test all the various methods proposed with a dataset that contains a more significant number of patients to prove the validity of these architectures. Another interesting future work could be to combine the predicted section of the same brain and create a 3D model of the ischemic stroke. Additionally, since in one of the predicted patient (Appendix A.1.10, A.2.10, A.3.10, B.1.10) the approaches generate false-positive leading to random classification and segmentation in isolated areas of the brain, a possible future research could be to find an algorithm to improve the efficiency of the post-processing technique to avoid these false-positive discoveries. Furthermore, the algorithm proposed to extract the skull from CTP images is not perfect; thus, it can be improved to test if it is lead to better results.

7

Conclusion

The thesis presented four different architectures, all based on a CNN structure, to predict in which part of the brain a stroke can occur and to classify or to segment the various regions, penumbra, and core, of a possible ischemic stroke. The input of the distinct CNNs is based on a dataset of CPT images over time from a group of 11 patients. A list of vectors made of a series of 16x16 tiles is extrapolated with a sliding window technique from these CTP images of the various brain sections created over a period of time. Each vector contains the same portion of the pixels of a specific brain section during the injection of an iodinated contrast agent inside cubital veins to enhance contrast in the tissue.

The output of each model is the same section of the brain included as the input, with the addition of manual annotation of the possible areas of the penumbra and core made by a specialist. This approach is suggested to have the possibility to classify each vector of tile extracted with a specific class, corresponding to one of the possible outcomes to recreate, at the end of the training and testing of the model, a predicted brain section image. The results obtained in this thesis work are very encouraging: the best test result afterK-Cross Fold Validation is found by using a Pixel by Pixel Segmentation approach (Chap. 5), learned with an SGD optimizer, the “Dataset 1”, and the Dice loss function.

77

Tomasetti Luca Chapter 7 Conclusion

The best overall result is 97.06% accuracy over all the four classes (background, brain, penumbra, core). The accuracy achieved with one of the proposed architecture should be the foundation for future researches: a more extensive and heterogeneous dataset, an improved version of the algorithm to extract the skull, a new version of the model containing a more significant number of hidden layers and an increment of the training time.

List of Figures

2.13 Pictorial representation of max pooling and average pooling . . . 22 2.14 Architecture of the U-Net. . . 24 3.7 Difference between the same brain section before and after the contrast

enhancement. . . 36 3.8 Brain section after the extraction of the different regions. . . 36 3.9 Example of the sliding window technique. . . 38 4.1 Focus of chapter four. . . 41 4.2 Overview of the input and output section for the CNN architectures. . . . 42 4.3 Example of post processing steps for the Tile Classification approach. . . 43 4.4 General structure for the Tile Classification architecture 1. . . 44 4.5 General structure for the Tile Classification architecture 2. . . 45 4.6 General structure for the Tile Classification architecture 3. . . 46 4.7 Different learning curves for the first architecture . . . 51 4.8 Different learning curves for the second architecture . . . 54 4.9 Different learning curves for the third architecture . . . 56 5.1 Focus of chapter five.. . . 61

79

Tomasetti Luca LIST OF FIGURES

5.2 Overview of the input and output section for the U-net architecture. . . . 62 5.3 Architecture of a U-Net. . . 63 5.4 Pixel by Pixel segmentation network structure. . . 65 5.5 Different learning curves for the U-Net architecture . . . 67 6.1 Focus of chapter six. . . 69

List of Tables

4.4 Example of brain section comparison for patient 2 with different techniques

of the first architecture. . . 49 4.5 Comparison of the statistical information for prediction on Patient 2 based

on two different datasets. . . 50 4.6 Confusion Matrix for Patient 2 with the normal dataset and SGD optimizer. 50 4.7 Confusion Matrix for Patient 2 with the normal dataset and Adam optimizer. 50 4.8 Confusion Matrix for Patient 2 with the augmented dataset and SGD

optimizer. . . 51 4.9 Confusion Matrix for Patient 2 with the augmented dataset and Adam

optimizer. . . 51 4.10 Example of brain section comparison for patient 2 with different techniques

of the second architecture. . . 52 4.11 Confusion Matrix for Patient 2 with the normal dataset and SGD optimizer. 53 4.12 Confusion Matrix for Patient 2 with the normal dataset and Adam optimizer. 53 4.13 Comparison of the statistical information for prediction on Patient 2 based

on two different datasets for the second architecture. . . 53 4.14 Confusion Matrix for Patient 2 with the augmented dataset and SGD

optimizer. . . 54 4.15 Confusion Matrix for Patient 2 with the augmented dataset and Adam

optimizer. . . 54 4.16 Example of brain section comparison for patient 2 with different techniques

of the third architecture. . . 55 4.17 Confusion Matrix for Patient 2 with the normal dataset and SGD optimizer. 55 4.18 Confusion Matrix for Patient 2 with the normal dataset and Adam optimizer. 55 4.19 Confusion Matrix for Patient 2 with the augmented dataset and SGD

optimizer. . . 55 4.20 Confusion Matrix for Patient 2 with the augmented dataset and Adam

optimizer. . . 55 4.21 Comparison of the statistical information for prediction on Patient 2 based

on two different datasets for the third architecture. . . 57 4.22 Accuracy & Loss for all models. . . 58 5.1 Layers summary of Pixel by Pixel Segmentation approach. . . 64

81

Tomasetti Luca LIST OF TABLES

5.2 Example of brain section comparison for patient 2 with different techniques on the U-net architecture. . . 66 5.3 Accuracy & Loss for all models related with the pixel by pixel segmentation

approach. . . 68 6.17 Example of brain section comparison for patient 8 with different techniques

on the U-net architecture. . . 73 6.18 Example of brain section comparison for patient 11 with different

tech-niques on the U-net architecture. . . 74