🔄 CNN to Inferno Pipeline

Example on CNN with Inferno Pipeline Example for a Predictions

Pipeline demonstration:

import os
from pathlib import Path
from rich.console import Console
from CNN import run_inferno_prediction, predict_from_image_path

# Move the working directory to the parent folder
WORKDIR = Path("/home/maxvill/InfernoCalibNet")
os.chdir(WORKDIR)

# Initialize console for rich printing
console = Console()

# Retrieving logits from CNN for a given image
(logit_efu, logit_ate), *_ = predict_from_image_path(
    image_path="/home/maxvill/InfernoCalibNet/data/raw/xrays/xrays6/00012236_010.png"
)

predictor_sets = [
    {
        "LOGIT_EFFUSION": 0.2818,
        "LOGIT_ATELECTASIS": -0.2889,
        "AGE": 48,
        "VP": "AP",
        "GENDER": "M",
    }
]

predictand_input = {"LABEL_EFFUSION": [0, 1], "LABEL_ATELECTASIS": [0, 1]}

# Running Infeno prediction
result = run_inferno_prediction(
    predictor_sets=predictor_sets, predictand_input=predictand_input
)

Prediction Result of CNN

      Label   Logit  Confidence  Prediction
   Effusion  0.2818      0.5700           1
Atelectasis -0.2889      0.4283           1

🚀Params for Inferno

{
    'input_csv': 'data/inferno/calibration_test.csv',
    'model_path': 'data/inferno/combinedML50/learnt.rds',
    'input_values': {
        'AGE': [48],
        'GENDER': ['M'],
        'VP': ['AP'],
        'LOGIT_EFFUSION': [0.2818],
        'LOGIT_ATELECTASIS': [-0.2889],
        'LABEL_EFFUSION': [0, 1],
        'LABEL_ATELECTASIS': [0, 1]
    },
    'predictors': ['AGE', 'GENDER', 'VP', 'LOGIT_EFFUSION', 'LOGIT_ATELECTASIS'],
    'predictands': ['LABEL_EFFUSION', 'LABEL_ATELECTASIS']
}

Registered doParallelSNOW with 10 workers

Closing connections to cores.

========================================================
Single Patient Clinical Decision Report
========================================================
Patient Predictor Data (Features Only):
  AGE GENDER VP LOGIT_EFFUSION LOGIT_ATELECTASIS
1  48      M AP         0.2818           -0.2889

Predicted Probabilities for Outcomes (%):
  Outcome Probability
1   E0 A0        25.0
2   E1 A0        36.2
3   E0 A1        20.0
4   E1 A1        18.8

Expected Utilities for Clinical Actions:
                        Action Expected_Utility
1             Send to Hospital              0.5
2     Start Drainage Treatment              0.4
3 Start Bronchodilator Therapy              0.6
4              Supportive Care              0.5
5              Observe Closely              0.5

Recommended Clinical Action:
Start Bronchodilator Therapy

🧠 Recommendation Check
✅ Prediction 1 is reliable (±17.3%).
⚠️ Prediction 2 shows moderate uncertainty (±25.8%). Consider reviewing.
✅ Prediction 3 is reliable (±19.5%).
✅ Prediction 4 is reliable (±13.2%).

🧠 What is display?: Inferno Prediction Workflow

This output illustrates how the Inferno framework generates calibrated probabilistic predictions for clinical outcomes by combining CNN derived logits with auxiliary patient data.

🔀 Workflow Overview

1. Image Processing & Logit Extraction A chest x-ray image is passed through a CNN, which outputs raw logits for conditions such as Effusion and Atelectasis. These serve as structured summaries of the image.

2. GradCAM Heatmap Generation A GradCAM visualisation is created to highlight regions most influential to the CNNs decision, aiding interpretability.

3. Predictor Set Construction Patient specific predictor sets are prepared combining:

   • CNN logits
   • Auxiliary data (e.g., age, gender, view position)

4. Inferno Input Specification Predictors and predictands are defined:

   {
       "input_csv": "data/inferno/calibration_test.csv",
       "model_path": "data/inferno/combinedML50/learnt.rds",
       "input_values": {
           "AGE": [48],
           "GENDER": ["M"],
           "VP": ["AP"],
           "LOGIT_EFFUSION": [0.2818],
           "LOGIT_ATELECTASIS": [-0.2889],
           "LABEL_EFFUSION": [0, 1],
           "LABEL_ATELECTASIS": [0, 1]
       },
       "predictors": ["AGE", "GENDER", "VP", "LOGIT_EFFUSION", "LOGIT_ATELECTASIS"],
       "predictands": ["LABEL_EFFUSION", "LABEL_ATELECTASIS"]
   }

5. Prediction Execution

📊 Interpreting the Output

Prediction Result of CNN

Label	Logit	Confidence	Prediction
Effusion	0.2818	0.5700	1
Atelectasis	-0.2889	0.4283	1

Predicted Probabilities for Outcomes (%)

Outcome	Probability
E0 A0	25.0
E1 A0	36.2
E0 A1	20.0
E1 A1	18.8

Expected Utilities for Clinical Actions

Action	Expected Utility
Send to Hospital	0.5
Start Drainage Treatment	0.4
Start Bronchodilator Therapy	0.6
Supportive Care	0.5
Observe Closely	0.5

🚀 Recommended Action: Start Bronchodilator Therapy

🧠 Recommendation Check

• ✅ Prediction 1 is reliable (±17.3%)
• ⚠️ Prediction 2 shows moderate uncertainty (±25.8%)
• ✅ Prediction 3 is reliable (±19.5%)
• ✅ Prediction 4 is reliable (±13.2%)

🧼 Probability Function

For each combination of predictors \(x\) and predictands \(y\), Inferno estimates the following probability:

\[ P(y \mid x, \text{data}) \]

Explicitly, for example:

\[ P(\text{label-effusion}, \text{label-atelectasis} \mid \text{logit-effusion}, \text{age}, \text{gender}, \text{vp}) \]

where:

• \(x\) includes CNN logits (e.g., logit-effusion) and optionally other auxiliary variables (e.g., age, gender, vp),
• \(y\) includes the labels to predict (e.g., label-effusion, label-atelectasis).

This structured output supports clinician interpretation by combining model predictions with uncertainty estimates and actionable recommendations.