Research map / Architectures and efficiency / CNN design

RLPeri: Accelerating Visual Perimetry Test with Reinforcement Learning and Convolutional Feature Extraction

AAAI 2024 · Architectures and efficiency › CNN design · approach: Dense matching and estimation CNNs · 3 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes RLPeri, a branching dueling Q-network agent that learns the sequence of test locations and initial stimulus values for visual perimetry, using potential-based reward shaping and a pair of 3D seen/not-seen state matrices with group-wise and pointwise convolutions, and reports about 10-20% fewer stimuli than ZEST and SORS at comparable MSE.

Problem

Visual perimetry tests are long and tiring for patients, and reducing examination time trades off against threshold estimation accuracy.

Approach

Formulates perimetry as an MDP in which an RL agent (BDQ-style branching dueling network) picks the next location and starting stimulus, with ZEST estimating thresholds; state is a pair of 3D seen/not-seen count matrices processed by group-wise and pointwise convolutions, and reward is shaped by a potential equal to negative MSE of discovered locations.

Key results

• RLPeri used 310.56 stimuli vs 348.15 for ZEST at sigma=1 (MSE 0.938 vs 0.949), about 10-15% faster than ZEST across stopping criteria.
• RLPeri is about 15-20% faster than SORS (e.g., 221.19 vs 279.29 stimuli at sigma=3).
• TOP and DS are faster (54 and 156.31 stimuli) but have much higher MSE (32.35 and 15.17), and PASS with all locations has MSE 15.03.

Measured effects reported in the paper

Surprising findings

• Using the number of stimuli as the reward did not consistently produce faster tests, which the authors attribute to neglecting initial stimulus recommendation.

Limitations

• Evaluation is done in a simulator built from the Rotterdam dataset with estimated FOS curves using a fixed standard deviation, not on real patients.
• MSE differences from ZEST and SORS are small, and SORS-only MSE is better at sigma=3, so accuracy is only comparable.
• Some baseline results (DS, TOP) are taken from prior paper rather than re-run.

Experimental setup

Datasets: Rotterdam visual field dataset

Benchmarks: Rotterdam visual field dataset (simulated perimetry)

Metrics: number of stimuli presented mean squared error

Baselines: ZEST SORS PASS DS TOP

Builds on: ZEST Branching Dueling Q-Network DQN potential-based reward shaping

Closest papers by meaning

• FOCUS: Benchmarking Retinal Model Generalization from Foundation Vision Encoders to Multimodal LLMs (NeurIPS 2026 · similarity 0.84)
• RL-SeqISP: Reinforcement Learning-Based Sequential Optimization for Image Signal Processing (AAAI 2024 · similarity 0.83)
• MedEyes: Learning Dynamic Visual Focus for Medical Progressive Diagnosis (AAAI 2026 · similarity 0.83)
• Improving Vision-language Models with Perception-centric Process Reward Models (CVPR 2026 · similarity 0.82)
• Statistical Test for Attention Maps in Vision Transformers (ICML 2024 · similarity 0.82)
• Like an Ophthalmologist: Dynamic Selection Driven Multi-View Learning for Diabetic Retinopathy Grading (AAAI 2025 · similarity 0.82)