 Hello everyone, my name is Prashasthi Kanika and I am presenting my research paper on ensemble model for dementia prediction using MRI. Under the guidance of Dr. Manoj Sankhe from NMIMS and Dr. Deepak Patkar from Nanavati Hospital, dementia is a syndrome in which there is deterioration in memory, thinking, behavior and the ability to perform everyday activities. According to World Health Organization, dementia cases are increasing rapidly and it is predicted that by 2050 the count can go up to 139 million. Hence, there is a need of an efficient method that can predict dementia at an early stage. The modality used for research is MRI. MRI has advantages over other modalities like absence of ionizing radiation, increased imaging flexibility and better tissue contrast. According to Bigler, brain volume is the parameter that helps in differentiating a demented subject from a controlled one. According to another study, brain volume loss associated with very early dementia can be detected by MRI even before cognitive symptoms appear. The right hand side picture shows difference in brain size for normal and severely demented case. The right segment looks smaller and shows shrinkage of brain. The applications of this study include supporting disease diagnosis, understanding mechanisms and tracking clinical progression of disease and in pharmaceutical research to compare the effects of new drugs on different populations of patients. This slide shows summary of some studies that show association between loss of brain volume and neurodegenerative diseases like dementia and mild cognitive impairment. Here, brain volume estimation approaches with their advantages and limitations are discussed. Method names are stereology, pixel counting and planimetric methods. Their limitations include issues related to resolution, bias classification and lack of accuracy and reproducibility. Hence, based on the literature review, we can infer that most of the existing methods are based on small sample size, require human interaction, time consuming and less accurate in terms of tissue classification. Hence, an accurate method is required for early detection of dementia. Here, the dataset considered for research are discussed. One source of data is Nanavati Hospital and the other source of data is OSS dataset. OSS is a publicly available dataset that can be used for research purposes. The type of data is cross-sectional, that means each patient is scanned only once. Classes of subjects are demented and non-demented and age range is from 14 to 98 years. Incomplete data means data with missing values is not considered for research. The proposed model can be divided into two simple modules. The first one helps us in estimation of brain volume whereas the second module helps in classification of dementia. So, let's discuss the first module for brain volume estimation. Now, let's discuss the proposed model. First three-dimensional volumetric images are taken as input. After that, the noise is removed and features are enhanced in image preprocessing step. In brain extraction, the brain voxels are extracted and non-brain voxels are rejected. In brain boundary adjustment module, the actual boundary of the brain gets modified by taking some useful voxels and removing some non-useful voxels and after that the brain volume is calculated. Let's understand the proposed model with the help of images. The first 3D image sequences are taken as input that is shown in first row. Then second row shows images after preprocessing with GMI filtering, third row shows images after brain extraction and last row shows images after volume refinement. The reference volume for this particular patient number 6 is 1333 cm3. If we don't apply our model, then in that case the volume value comes to 1162.16 and accuracy was 87.16%. When we apply our proposed model with gradient magnitude image filtering, then the volume value comes to 1332.52 cm3 and the accuracy becomes 99.93%. So this way at 12.77% increment in accuracy is observed when we apply our proposed model with gradient magnitude image filtering. This slide shows two images. The left hand side one is before refinement and right hand side image is after refinement. There are some extra voxels peaks in left hand side image that are suppressed in right hand side image. Like this we get around 190 to 250 slices per patient and volume refinement is applied on all slices in parallel. Now here the proposed model with gradient magnitude recursive Gaussian filtering results are discussed. The first row shows the input image sequence. The second row shows the images after applying GMRG filtering, third row shows the results after extraction of brain and last row shows the results after volume refinement. Now the reference volume was 1333 because we are talking about patient number 6 only here. The volume and accuracy values were earlier 1162.16 and 87.16% respectively and when we apply our proposed model with GMRG filtering, then the newly calculated volume value comes to 1350.7 centimeter cube and accuracy becomes now 98.67%. So this way 11.51% increment in accuracy is observed when we use our proposed model with GMRG filtering. Now same model is tested with GAD filtering that is gradient and isotropic diffusion filtering. So first row is input images, second row is after applying the GAD filtering, third row is after brain extraction and last row is after volume enhancement. Now reference volume without using our model and accuracy without using our model are same as earlier slide but when we apply our proposed model with GAD filtering the volume value now become 1150.88 centimeter cube and accuracy after using our model becomes 86.31%. So there is no increment in accuracy rather than the accuracy value is reduced that means the proposed model with GAD filtering is not giving good results for my data set. Now this slide shows the comparative analysis using both the data sets means OSS data and nano-VT data. Now in this graph the red light shows GMI of accuracy, the green light shows the GMRG accuracy, purple line shows gradient and isotropic diffusion accuracy and the blue line shows the accuracy without the proposed model. So in first case with OSS data it can be clearly seen that the GMI filtering performs the best and at second level GMRG filtering proposed model performs but in case of nano-VT data the proposed model with GMI and GMRG filtering both are giving almost equal values. Now here the comparative analysis using all the variables are shown. So in first case the results are shown using OSS data and in second graph the results are shown using nano-VT hospital data. On OSS data set the proposed model performs best with GMI filter as compared to other filters with an average accuracy of 94.156%. On right hand side graph on nano-VT data set the proposed model gives comparable average accuracy using GMI and GMRG filters as 93.65% and 93.40% respectively. That means there is very minor difference between the accuracy obtained through both the approaches. Now in first module we have seen how the brain volume is calculated. Now the second volume discusses which method best classifies the dementia for the given data set. Now here some of the approaches used for dementia classification are listed. Through the literature review it was informed that the sample size was small and the accuracy was low. Now when we applied variety of machine learning algorithms on all data sets then these are the results. In first column model name is listed. In second column model type is listed. In third column accuracy percentage by four fold cross validation are discussed. Now by four fold cross validation we mean that the 75% of the data values are used for training the model and 25% of the data values are used for testing the model. Now fourth column shows the accuracy percentage by five fold cross validation. That means 80% of the data items are used for training the machine learning model and 20% of the values are used for testing the model. And the last column shows the accuracy percentage by 10 fold cross validation. That means 90% of the data items are used for training the machine learning model and 10% of the values are used for testing. Now here we can see that ensemble model with back trees performs the best with an accuracy of 97.8%. Now conclusion, on OSS dataset the proposed model performs best with GMI filter as compared to other filters with an average accuracy of 94.156%. The proposed model with GMI and GMRG filtering increase volume estimation accuracies by 7.68% and 8.04% respectively. On Nanauti dataset the proposed model gives comparable average accuracies using GMI and GMRG filters as 93.65% and 93.40% respectively. Proposed models with GMI and GMRG filtering increase the volume estimation accuracies by 10.41% and 10.16% respectively. For dementia prediction, ensemble model with back trees outperforms other methods with classification accuracy of 97.8%. Now here are some references, thank you. My final thesis presentation is yet to be happened that's why I have not disclosed all my results in detail so in case of any query or research related collaboration please connect with me on my given mail id. Thank you once again.