The human hippocampal formation is a very small anatomical region in the brain, yet it is essential in the declarative memory system, which encompasses stored short-term episodic memory. The accelerated neuronal loss in the hippocampus region compared to normal elderly has been shown as one of the earliest detectable signs of many neurodegenerative diseases, such as Alzheimer’s disease (AD). Measures of change in hippocampal volume derived from longitudinal magnetic resonance imaging (MRI) are a well-studied biomarker of disease progression in AD and are used in clinical trials to track the therapeutic efficacy of disease-modifying treatments. However, conventional longitudinal MRI change measures are derived using a computer algorithm, called deformable image registration, which can be easily confounded by noise in MRI scans, such as head motion or MRI artifacts, resulting in over-estimation or under-estimation of hippocampal atrophy. This dissertation introduces novel approaches to quantify disease progression in the hippocampal region with deep learning techniques designed to reduce the influence of noise in MRI scans. The underlying assumption of these methods is that a deep learning network trained to derive temporal information from longitudinal scan pairs (such as which scan was acquired first, or how long the interval between two scans is) implicitly learns to detect progressive changes that are related to aging and disease progression, as opposed to changes related to MRI noise. In the first such network, called DeepAtrophy, a single scalar measure was derived to infer the relative disease progression rate from a pair of longitudinal MRI scans, which is compared to the progression rate of the normal aging group. The results of DeepAtrophy on T1-weighted MRI show 89.3% accuracy in inferring scan temporal information on held-out test data (compared to 76.6% in conventional methods), and the scalar disease progression measure derived from DeepAtrophy is significantly different between cognitively unimpaired individuals with and without evidence of preclinical Alzheimer’s disease pathology. In contrast, conventional methods fail to detect any significant difference between the preclinical AD group and the cognitively unimpaired group. To understand how the model made decisions from longitudinal scan pairs, I developed a second model, Regional Deep Atrophy (RDA), which generates a spatial map identifying regions in the image that contribute to the final estimation of disease progression, without compromising sensitivity to longitudinal changes. Analysis of these maps suggests that areas most related to atrophy in the hippocampal region are the hippocampus and peripheral gray and white matter regions, while areas most related to the expansion of tissue are ventricles and CSF. Both algorithms were compared to the state-of-the-art conventional algorithm ALOHA, which is based on deformable image registration, showing large improvements in the ability of deep learning algorithms to infer temporal information from scan pairs, and comparable ability to detect differences in rates of atrophy between groups with different degrees of Alzheimer’s pathology. Further investigation focused on comparing T1-weighted and T2-weighted MRI in the context of detecting atrophy in the hippocampal region and showed that the improvements of DeepAtrophy over ALOHA extend to T2-weighted MRI but that T1-weighted and T2-weighted MRI mostly have similar performance in temporal inference and group discrimination tasks. Overall, DeepAtrophy and RDA show excellent temporal inference accuracy and are able to detect increased disease progression in individuals with preclinical Alzheimer’s disease, demonstrating a potential to be a more accurate and trustworthy surrogate biomarker to estimate disease progression in future clinical trials.