Predicting treatment response of tumor spheroids from radiomics analysis of post-treatment dynamics

contact: pejman.shojaee@htw-dresden.de

Affiliation: HTW Dresden – University of Applied Sciences

Background

This repository contains all scripts and workflows used in the study submitted to GigaScience. The project implements an AI-driven radiomics pipeline for predicting long-term relapse of tumor spheroids from early post-treatment brightfield image data.

It contains:

• MLTimeSeriesModel.py – a configurable ML pipeline with variance filtering, scaling, SMOTE, multiple feature-selection strategies, and a zoo of classifiers; includes cross-validation with 95% CIs, bootstrapped evaluation, probability-based ROC utilities, and global seed helpers for reproducibility.
• Visualization.py – a plotting suite for confusion matrices, ROC curves (with optimal threshold), probability distributions, accuracy vs. Day-of-Relapse (DoR) with weighted exponential fits, and stacked SCP (Share of Controlled Proportion) bar charts for treatment arms (dose, temperature, time).
• Manuscript_codes_fixed2.ipynb – an analysis notebook (used in the manuscript).

While radiomics is routinely used in patient-level imaging, it has not yet been applied to three-dimensional (3D) multicellular tumor spheroids — a physiologically relevant in vitro model that recapitulates gradients of oxygen, nutrients, and therapy resistance observed in real tumors.

Here, we present an end-to-end machine learning framework that predicts whether individual spheroids relapse or remain controlled, based solely on early post-treatment imaging features. Our workflow integrates radiomic feature extraction, feature selection, and classification models to infer long-term treatment outcomes.

📦 Installation

# Python 3.10+ recommended
pip install -U numpy pandas scikit-learn imbalanced-learn xgboost matplotlib seaborn statsmodels
# Optional (for deterministic seeding)
pip install torch

Notes:

• All CSVs in data/ are exactly those used in the manuscript models (training and test).
• Splits used in the paper are under data/splits/.

Quick start

You can easily use the uploaded time range data and use it to visualize the results. You can also use your own data and the pipeline to calculate the classified target, such as relapsed and controlled cases.

from MLTimeSeriesModel import MLTimeSeriesModel
from Visualization import Visualization
import pandas as pd

# Load your dataset
df = pd.read_csv("data/short_final_with_updated_diagnosis.csv")
X = df.drop(columns=["diagnosis"])
y = df["diagnosis"]

# Initialize pipeline
model = MLTimeSeriesModel()
model.select_feature_selection("SelectKBest_f_classif")
model.select_model("RandomForest")
model.build_pipeline()

# Cross-validation
results = model.cross_validate(X, y)
print(results["mean_auc"], results["auc_confidence_interval"])

# Visualization example
viz = Visualization(output_dir="results", y_true=y, y_pred=None)
viz.plot_confusion_matrix()

maintainers

• Pejman Shojaee (@pejmanshojam31) — lead
• Tom Bischopink (@tomb556)