Welcome to InfernoCalibNet
Uncertainty aware predictions for medical AI using CNN and Bayesian nonparametrics framework (Inferno)
Through uncertainty aware modeling, this research explores the application of Bayesian regression and convolutional neural networks (CNNs) to assist medical decision making. The CNN serves as a feature extractor in this configuration, converting complex visual input into a vector of real valued class values. These ratings are treated as structured summaries of the visual information rather than as final judgments.
This project uses Inferno, a Bayesian method that turns the CNNs raw confidence scores (logits) into well adjusted probability estimates. By adjusting for patient specific base rates and incorporating previous data, this statistical module allows clinicians to make decisions based on presented benefit rather than strict classification. Because the final decision is visible, individualized and based on Bayesian reasoning, the design eliminates the need to “explain” the CNN itself by separating prediction from action.
Through a series of hypothetical experiments created to mirror the needs of personalized medicine, the project investigates Inferno’s durability and medicinal usefulness. GradCAM heatmaps, which visually highlight where the CNN focuses its attention based on features from the base layer of the network, are included alongside the calibrated outputs. The combined probabilistic and location outputs are meant to help clinicians go beyond classification and provide informed, personalized medicine by helping them customize treatment choices for specific patients.
🧠 System Architecture
| Maksim Ohvrill |
 
|
| PierGianLuca Porta Mana |
|
📖 Citation
If you use this software or refer to its documentation, please consider citing:
@software{infernoCalibNet,
author = {Maksim Ohvrill and PierGianLuca Porta Mana},
title = {InfernoCalibNet: Bayesian Calibration of CNN Outputs to Aid Clinical Decision-Making},
year = {2024},
version = {1.0},
date = {2024-04-29},
url = {https://m4siko.github.io/InfernoCalibNet/}
}