[Paper Review] A multi-task deep learning model for the classification of Age-related Macular Degeneration

Qingyu Chen, Yifan Peng+6 more|arXiv (Cornell University)|Dec 2, 2018

Retinal Imaging and Analysis15 references46 citations

TL;DR

The paper introduces a multi-task deep learning model that classifies AMD severity by first scoring four AMD characteristics (drusen area, geographic atrophy, increased pigment, depigmentation) and then computing the AREDS 1-9 severity score, achieving superior accuracy and F1-scores versus the state-of-the-art on AREDS and AREDS2 datasets.

ABSTRACT

Age-related Macular Degeneration (AMD) is a leading cause of blindness. Although the Age-Related Eye Disease Study group previously developed a 9-step AMD severity scale for manual classification of AMD severity from color fundus images, manual grading of images is time-consuming and expensive. Built on our previous work DeepSeeNet, we developed a novel deep learning model for automated classification of images into the 9-step scale. Instead of predicting the 9-step score directly, our approach simulates the reading center grading process. It first detects four AMD characteristics (drusen area, geographic atrophy, increased pigment, and depigmentation), then combines these to derive the overall 9-step score. Importantly, we applied multi-task learning techniques, which allowed us to train classification of the four characteristics in parallel, share representation, and prevent overfitting. Evaluation on two image datasets showed that the accuracy of the model exceeded the current state-of-the-art model by > 10%.

Motivation & Objective

• Motivate automated, scalable AMD severity grading beyond manual 9-step scoring due to time and cost of expert grading.
• Mimic human reading-center grading by predicting four AMD characteristics and deriving the 9-step AREDS severity scale.
• Leverage multi-task learning to share representations across characteristics and reduce overfitting.
• Evaluate robustness and generalizability on AREDS and AREDS2 datasets and compare to existing models.

Proposed method

• Propose a multi-task architecture where four classifiers share initial layers (Inception-V3 based feature extractor) and have task-specific heads for each AMD characteristic.
• Use pre-processing steps to normalize brightness, crop to square, and resize to 512x512 pixels.
• Apply data augmentation (rotation, flips) and dropout to mitigate overfitting; train with Adam optimizer using cross-entropy losses.
• Implement a two-phase training: (i) train all tasks in parallel to learn shared representations, (ii) fine-tune task-specific heads while freezing shared layers.
• Ground predictions of four characteristics to compute the AREDS severity score per the defined mapping of drusen area, geographic atrophy, increased pigment, and depigmentation (per Table 1 in the paper).
• Compare against Grassmann et al.’s model and a CNN baseline trained on AREDS, using five-fold cross-validation on AREDS and independent testing on AREDS2.

Experimental results

Research questions

• RQ1 Does a multi-task model outperform single-task and baseline CNN models in predicting the AREDS 1-9 severity scale on AREDS and AREDS2?
• RQ2 Do intermediate predictions of four AMD characteristics improve final severity scoring and interpretability?
• RQ3 How does transfer learning from ImageNet and multi-task training affect generalization across AREDS and AREDS2 datasets?
• RQ4 Which AMD characteristic drives most of the error in severity scoring, and how does class balance affect performance?
• RQ5 Is the multi-task approach robust to distribution shifts between AREDS and AREDS2 datasets?

Key findings

• The multi-task model outperformed the CNN baseline and Grassmann model across metrics on both AREDS and AREDS2 datasets.
• On AREDS2, the multi-task model achieved approximately 5% higher weighted F1-score, about 3% higher Kappa, and ~4% higher accuracy than the CNN baseline.
• Compared to single-task learning, multi-task learning yielded about 4% higher weighted F1-score and ~2% higher other metrics.
• Transfer learning with pre-trained weights from ImageNet contributed to improved generalization, especially on AREDS2.
• Error analysis showed drusen area classification as the main bottleneck, with lower per-class accuracy, suggesting data imbalance and limited samples for some classes.