File size: 4,362 Bytes
4a48fa2
 
 
 
 
 
 
 
efdc633
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4a48fa2
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
---
license: gpl-3.0
language:
- en
metrics:
- f1
- accuracy
---
# HCAT-FusionNet:  Multimodal Preprocessing and Fusion for Survival and Recurrence Prediction,using variational autoencoders and cross-modal attention for holistic healthcare outcome prediction.

This repository contains preprocessing pipelines and training framework for the **HANCOCK (Head and Neck Cancer Cohort)** dataset used in the **Hancothon25 Challenge** at MICCAI 2025. The challenge focuses on predicting **5-year survival** and **2-year recurrence** using multimodal patient data (clinical, pathological, semantic text, spatial histopathology, temporal blood tests).

Our solution introduces **novel preprocessing, imputation, and fusion strategies** to extract robust 512-dimensional embeddings from heterogeneous modalities, followed by **advanced multi-modal training**.

---

## Features

* **Clinical Data Preprocessing**: Advanced imputation ensemble + VAE-based handling of missing data.
* **Pathological Data Preprocessing**: Probabilistic imputation with graph smoothing and 512-d embeddings.
* **Semantic Text Processing**: ClinicalBERT / TF-IDF + SVD pipelines for histories, reports, and surgery descriptions.
* **Spatial Histopathology Aggregation**: Transformer-based aggregation of patch-level features with spatial awareness.
* **Temporal Blood Data**: Physiology-aware normalization, KNN refinement, and LSTM encoder for sequential signals.
* **Fusion Training**: Multi-modal VAE with attention-based cross-modal imputation, joint latent space learning, and uncertainty quantification.
* **Evaluation**: Binary classification of survival and recurrence, reporting accuracy and F1-score.

---



## Documentation

Detailed explanations of each pipeline are available here:

* [Clinical Preprocessing](https://github.com/Ragu-123/hcat-fusionnet/blob/main/Documentation/clinical.md)
* [Pathological Preprocessing](https://github.com/Ragu-123/hcat-fusionnet/blob/main/Documentation/pathological.md)
* [Semantic Text Processing](https://github.com/Ragu-123/hcat-fusionnet/blob/main/Documentation/semantic.md)
* [Spatial Histopathology](https://github.com/Ragu-123/hcat-fusionnet/blob/main/Documentation/spatial.md)
* [Temporal Blood Data](https://github.com/Ragu-123/hcat-fusionnet/blob/main/Documentation/temporal.md)
* [Training Pipeline](https://github.com/Ragu-123/hcat-fusionnet/blob/main/Documentation/train.md)

---

## Results

From the enhanced training pipeline (`train2.py`), the system achieved:

* **5-year Survival F1-score**: **0.80**
* **2-year Recurrence F1-score**: **0.95**
* **Average F1-score**: **0.875**

*(See [`enhanced_hcat_training_summary.json`](enhanced_hcat_training_summary.json) for full training logs and config.)*

---

## Methods Summary

* **Imputation**: Multi-modal VAE, KNN, PCA, and graph smoothing.
* **Embeddings**: Standardized 512-d representations across modalities.
* **Fusion**: Attention-based cross-modal integration with uncertainty weighting.
* **Classification**: Binary prediction of survival and recurrence with robust evaluation.

---

## Challenge Context

This work addresses the **HANCOCK multimodal dataset** provided for **Hancothon25 (MICCAI 2025)**. The dataset includes **763 patients** with modalities:

* Clinical structured data
* Pathology structured data
* Histopathology WSIs & TMAs
* Tabular blood test data
* Free-text clinical/surgery reports

Our framework is designed for **precision oncology**, enabling predictive modeling for treatment planning and follow-up.

---

## Performance & Insights

* The system demonstrates strong **generalization across modalities**.
* Temporal and pathological modalities improved recurrence prediction.
* Clinical and semantic features boosted survival classification.
* Fusion strategies with uncertainty modeling ensured robustness under missing modalities.

---

## Citation

If you use this repository, please cite:

```
@inproceedings{hcat_fusionnet_2025,
  title={HCAT-FusionNet: Multimodal Preprocessing and Fusion for Survival and Recurrence Prediction},
  author={Ragunath R, Sanjay S, Harish G},
  booktitle={MICCAI Hancothon25 Challenge},
  year={2025}
}
```

---

## Contact

For questions or collaborations:

* **Author**: [Harish G](https://github.com/Harish2404lll), [Ragunath R](https://github.com/Ragu-123), [Sanjay S](https://github.com/22002102)