Background:
The past decade has seen a fundamental shift in preclinical disease modelling toward New Approach Methodologies that reduce or replace animal experimentation. Brain organoids-three-dimensional, self-organising neural structures derived from human iPSCs — sit at the forefront of this shift. They recapitulate key developmental stages of the human cortex in vitro, including the formation of ventricular zones, cortical layers, and disease-specific structural abnormalities, with no analogous animal model capable of faithfully reproducing the human neurodevelopmental programme. Brain organoids represent one of the most promising New Approach Methodologies (NAMs) for studying human neurodevelopmental disorders without animal models or invasive human tissue access. However, their utility in high-throughput applications is limited by the need for molecular assays (immunofluorescence, scRNA-seq) to confirm disease phenotype. Automated classification of organoid disease status from routine brightfield microscopy, a non-invasive, label-free, scalable imaging modality — remains an unsolved computational challenge.

 

Hypothesis/Objective:
To develop and evaluate a hybrid quantum-classical machine learning pipeline capable of automatically classifying brain organoid disease status directly from routine brightfield microscopy images, without requiring molecular assays such as immunofluorescence or scRNA-seq.

A hybrid architecture combining classical deep learning feature extraction (ResNet18) with quantum computing components (quanvolutional filters and a variational quantum layer) can detect disease-relevant morphological signals in label-free brightfield images of brain organoids, enabling statistically significant discrimination between healthy and diseased organoid phenotypes.

 

Methods:
We present a hybrid quantum-classical pipeline applied to this task for the first time, evaluated on the publicly available Zenodo 10301912 dataset of 1,407 brightfield images from four clones: wt2D (healthy control) and TH2-7, A1A-1, B2A-2 (disease models of tyrosine-hydroxylase deficiency and tubulinopathy). The architecture combines ResNet18 pretrained feature extraction [1] with the Ingenii Quantum Hub SDK [2] — providing QuantumFilters2D (quanvolutional preprocessing) and circuit visualisation — and a PennyLane [3] variational quantum FC layer. A ZZ-feature-map quantum kernel SVM [4] provides an additional classification pathway. A two-stage training strategy with WeightedRandomSampler addresses the inherent 1:3 healthy:diseased class imbalance.

 

Results:
The Hybrid model achieves AUC = 0.801 at default threshold 0.5, yet balanced accuracy = 0.750 at the same threshold. At Youden-optimal threshold (t = 0.26), balanced accuracy rises to 0.824 with no change in AUC. This dissociation is a recurring challenge in imbalanced NAM classification tasks and is directly addressed by the threshold optimisation step. That is, the Ingenii Hybrid model achieves AUC = 0.801 (95% CI: 0.718–0.884) and balanced accuracy = 0.824 at the Youden-optimal threshold [5], with 87.5% healthy recall and 77.3% diseased recall on a 30-sample held-out test set. This is the first statistically significant disease classifier reported on this dataset, establishing a computational baseline for future quantum and classical NAM-enabling tools.

This study evaluated hybrid quantum-classical machine learning models for classifying brain organoids as healthy or diseased. As shown in Figure 1, the dataset comprised 120 images across four clones (wt2D, TH2-7, A1A-1, B2A-2), though it was notably imbalanced (30 healthy vs. 90 diseased), split evenly between two labs with imaging concentrated between days 10–30. This imbalance is a critical consideration when interpreting all subsequent results.

Feature extraction via ResNet18, visualized in Figure 3, revealed only moderate class separability in PCA space, with healthy and diseased organoids intermingling substantially across clone clusters-establishing a likely ceiling on model performance. Complementing this, Figure 4 shows that the quantum kernel matrices produced strongly diagonal structures in both the training (45×45) and test (15×15) kernels, indicating highly individualized sample representations with limited inter-sample similarity and further constraining the kernel SVM’s discriminative power.

Among the four models tested, Table 1 reveals that the Q-Kernel and Classical SVMs performed identically (AUC = 0.750, Balanced Accuracy = 0.732), suggesting the quantum kernel offered no advantage in isolation. Both Hybrid models improved AUC to 0.801, with the Hybrid @Youden threshold emerging as the strongest overall performer-achieving the best Balanced Accuracy (0.824) and F1 (0.850). The Hybrid @0.5 variant maximized healthy recall (1.000) but at the cost of F1 (0.667), making it less reliable overall. Notably, tighter bootstrap confidence intervals for Hybrid models (±0.083 vs. ±0.145–0.151) indicate greater stability. These findings are further corroborated visually in Figure 6, where Hybrid @Youden consistently leads across Balanced Accuracy, Specificity, and F1, while Q-Kernel and Classical SVMs perform nearly identically across all bars.

At the classification level, Table 2 shows that at the Youden threshold (t = 0.26), the best model correctly identified 7/8 healthy organoids (87.5%) and 17/22 diseased organoids (77.3%), with 5 false negatives and only 1 false positive-the missed diseased cases representing the greater clinical concern. This performance is placed in broader context by Figure 5(A), where the Hybrid model’s ROC curve achieves the highest AUC (~0.80), clearly surpassing both the Classical SVM and Q-Kernel baselines, with all models well exceeding the chance baseline (AUC = 0.50). Figure 5(B) reinforces this through a radar chart, confirming Hybrid @Youden as the most balanced model across all five metrics, while the Classical SVM shows a notable dip in specificity and the Q-Kernel dips in sensitivity.

Training dynamics shown in Figure 2 revealed mild overfitting, with validation balanced accuracy plateauing between 0.3–0.5 despite steadily decreasing training loss across ~50 epochs. The unfreezing of layers 3+4 caused a brief loss spike before recovery, and the persistent gap between training and validation accuracy reflects the difficulty of generalizing from a small, imbalanced dataset.

 

Conclusion:
Hybrid quantum-classical AI can extract disease-relevant morphological signals from brain organoid brightfield images, supporting the vision of organoids as fully autonomous, self-reporting in vitro disease models that require no molecular confirmation of phenotype.

 

Image/Figure:
Click to view full size
Click to view full size

Image/Figure Caption:
All Figures

 

Author
Bamidele Omotayo Awojoyogbe, PhD, MSc, BSc
Federal University of Technology Minna