Add RetinaUNet model with synthetic example and unit tests

[mtuann]

Contributor Information

• Name: Tuan Nguyen
• NetID: tuanmn2
• Email: tuanmn2@illinois.edu

Type of Contribution

• Model contribution

Original Paper

• Retina U-Net: Embarrassingly Simple Exploitation of Segmentation Supervision for Medical Object Detection
  https://proceedings.mlr.press/v116/jaeger20a/jaeger20a.pdf

Motivation

This PR adds a Retina U-Net style medical object detection model to PyHealth.
My goal here is to contribute a reusable PyHealth model implementation that
captures the main architectural idea of the original paper: dense object
detection with auxiliary segmentation supervision. I intentionally kept the
scope model-centric and dataset-agnostic on the PyHealth side. LIDC-specific
download, preprocessing, and larger reproduction experiments stay in a separate
companion project so that this PR remains focused and easier to review.

Relevance To PyHealth

• Healthcare relevance: Retina U-Net is a medical imaging detection model
  derived from a peer-reviewed healthcare AI paper.
• Package relevance: this adds a reusable object detection model to
  pyhealth.models without tying the contribution to one local dataset layout
  or one machine-specific workflow.

High-Level Description

This PR adds:

• a RetinaUNet model under pyhealth.models
• dense classification and box regression heads for detection
• an auxiliary segmentation decoder
• lightweight synthetic tests covering initialization, forward pass, backward
  pass, embedding mode, and inference mode
• a lightweight synthetic ablation example showing how to compare small model
  configurations without requiring any real dataset
• model API documentation and index updates

The example and tests are designed to be fast and self-contained. They do not
depend on local LIDC data and use synthetic tensors only. The main artifact for
review is the reusable RetinaUNet model implementation; the example and tests
exist to make the contribution easy to validate.

Files To Review

Main implementation:

• pyhealth/models/retina_unet.py

Model registration:

• pyhealth/models/__init__.py

Tests:

• tests/core/test_retina_unet.py

Example:

• examples/lidc_nodule_detection_retina_unet.py

Docs:

• docs/api/models.rst
• docs/api/models/pyhealth.models.RetinaUNet.rst

Testing Performed

Validated in a clean uv environment with:

• import probe from pyhealth.models
• python -m unittest tests.core.test_retina_unet -v
• a small synthetic ablation run from
  examples/lidc_nodule_detection_retina_unet.py

Observed validation outcome:

• RetinaUNet import succeeds
• unit tests pass
• the synthetic ablation example runs end-to-end and prints a configuration
  comparison summary

Relation To The Reproduction Project

This PR comes out of a CS598 DL4H reproducibility project based on the Retina
U-Net paper. In the companion project repository, I integrated the model into
an LIDC-IDRI pipeline with reproducible preprocessing and controlled
experiments. That external project is useful validation context, but it is not
the contribution proposed here. The actual contribution in this PR is the
PyHealth model implementation itself, together with its tests,
documentation, and lightweight synthetic ablation example.

Scope Limitations

• this is not an exact reproduction of the original paper benchmark
• the LIDC data pipeline is intentionally kept out of this PR
• the companion project currently evaluates detections with a slice-level
  evaluator rather than full 3D lesion-level matching
• this PR is intended to contribute a reusable model, tests, docs, and a
  minimal example rather than a full dataset-specific benchmark pipeline

Notes For Reviewers

• the example is synthetic on purpose so it stays lightweight and runnable
• the model contribution is the primary artifact of this PR
• no machine-specific paths, generated data, checkpoints, or logs are intended
  to be part of this pull request