Towards Alzheimer's disease classification through transfer learning

Abstract

Detection of Alzheimer's Disease (AD) from neuroimaging data such as MRI through machine learning have been a subject of intense research in recent years. Recent success of deep learning in computer vision have progressed such research further. However, common limitations with such algorithms are reliance on a large number of training images, and requirement of careful optimization of the architecture of deep networks. In this paper, we attempt solving these issues with transfer learning, where state-of-the-art architectures such as VGG and Inception are initialized with pre-trained weights from large benchmark datasets consisting of natural images, and the fully-connected layer is re-trained with only a small number of MRI images. We employ image entropy to select the most informative slices for training. Through experimentation on the OASIS MRI dataset, we show that with training size almost 10 times smaller than the state-of-the-art, we reach comparable or even better performance than current deep-learning based methods.

Not-so-supervised: A survey of semi-supervised, multi-instance, and transfer learning in medical image analysis

“…Task Domain Transfer type Brain Zhang and Shen (2012) MCI conversion prediction different same feature, multi-task Wang et al (2013) tissue, lesion segmentation same different instance, weight van Opbroek et al (2015a) tissue, lesion segmentation same different instance, weight Guerrero et al (2014) AD classification same different instance, align van Opbroek et al (2015b) tissue, lesion segmentation same different instance, weight Cheng et al (2015) MCI conversion prediction different same feature, multi-task Goetz et al (2016) tumor segmentation same different instance, weight Wachinger and Reuter (2016) AD classification same different instance, weight Cheplygina et al (2016a) tissue segmentation same different instance, weight Ghafoorian et al (2017) lesion segmentation same different feature, pretraining Kamnitsas et al (2017) segmentation of abnormalities same different feature, pretraining Alex et al (2017) lesion segmentation different same feature, pretraining Hofer et al (2017) AD classification same different instance, align Hon and Khan (2017) AD classification different different feature, pretraining Kouw et al (2017) tissue segmentation same, different instance, align Breast Huynh and Giger (2016) tumor detection different different feature, pretraining Samala et al (2016) mass detection same different feature, pretraining Kisilev et al (2016) lesion detection, description in mammography or ultrasound different same feature, multi-task Bi et al (2008) abnormality classification different same feature, multi-task Schlegl et al (2014) lung tissue classification different same/different feature, pretraining Bar et al (2015) chest pathology detection different different feature, pretraining Ciompi et al (2015) nodule classification different different feature, pretraining Shen et al (2016) lung cancer malignancy prediction different same feature, multi-task Chen et al (2017b) attribute classification in nodules different same feature, multi-task Hussein et al (2017) attribute regression, malignancy prediction different same feature, multi-task Cheplygina et al (2017) COPD classification same different instance, weight…”

Generalization Performance of Deep Learning Models in Neurodegenerative Disease Classification

“…Again, they did not perform subject level division. Lastly, Hon et al [24] utilized two stateof-the-art architectures, namely VGG16 and Inception V4 to classify AD. They used 5-fold cross-validation to obtain the results, with an 80% -20% split between training and test.…”

“…Throughout the work, we realized that a common misconception occurs in many different papers which use machine learning algorithms in 3D medical imaging. Performance of the models was often determined by dividing the pooled slices into training and test sets [21], [24]- [26], [42] (see Table I). Thus, training and test sets included the different brain slices of the same subjects.…”

Using transfer learning for classification of gait pathologies

“…Recently, the use of deep CNNs, such as VGG-16 [19] or pose based temporal-spatial networks [20], have significantly improved the performance of silhouette based gait recognition systems. Similar improvements have also been seen in the medical domain, especially in detecting Alzehimer's disease [21]. Thus, it can be expected that the use of such deep learning techniques will also improve the performance of the gait pathology classification systems.…”