01
The Bottleneck
ConvNeXtBase → 8-Qubit Circuit
Information bottleneck. 1024 features crushed into 8 qubits, collapsing any expressive structure downstream.
Failed Research · Quantum ML
Quantum vs. classical CNNs
for medical imaging.
A comparative study asking whether Hybrid Quantum-Classical networks can outperform a standard CNN in low-data melanoma classification. Four experiments, three dead ends, one verified quantum advantage.
+6.36%
Accuracy
vs. classical
+6.71%
Precision
vs. classical
+4.11%
F1-Score
vs. classical
4
Experiments
3 failed · 1 worked
01
The Question
Research Question
Does adding a quantum processing layer improve deep learning performance for melanoma detection?
Tested multiple Quanvolutional architectures against identical classical controls. After three failed experiments, Experiment 4 demonstrated a measurable quantum advantage on data-scarce melanoma classification.
02
The Journey
Four experiments ran end to end. Each node on the timeline records what was tried, what broke, and what survived.
02
Quanvolutional Layer
Quantum filter as feature extractor
Multiple iterations either overfit aggressively or matched the classical control point for point. No usable signal.
Failed 03
Recall-Focused
Simplified circuit · 9,942 test images
High recall (79%) traded for poor precision (42%). The output bottleneck pushed the model toward a single class.
Partial 04
High-Resolution
ResNet18 + Data Re-uploading + Projective Measurement
Quantum advantage verified: +6.4% accuracy, +6.7% precision, better generalization on the held-out set.
Success 03
Winning Architecture
The four moves that made Experiment 4 actually work.
01
ResNet18
Partially unfrozen classical backbone.
02
14 × 14 grid
196 feature patches — 4× the previous resolution.
03
Data Re-uploading
Features encoded twice to inject non-linearity into the circuit.
04
256-D Measurement
Projective output, up from 8 dimensions.
04
Results
Head to head on the held-out test set. Same data, same backbone — only the quantum layer differs.
Metric
QCNN
Classical
Δ
Accuracy
77.73%
71.37%
+6.36%
Precision
49.97%
43.26%
+6.71%
Recall
85.04%
91.95%
-6.91%
F1-Score
62.95%
58.84%
+4.11%
AUC
0.8906
0.8858
+0.5%
Finding
QCNN wins four of five metrics.
Classical edges recall only because it over-predicts the positive class. QCNN actually learns to discriminate — accuracy, precision, F1, and AUC all rise together.
05
Key Findings
01
Superior Generalization
Anti-overfitting effect
The classical model drove training loss down to 0.06 while validation loss climbed to 0.79. The QCNN stayed balanced, acting as a natural regularizer on a small dataset.
02
Escaping the Recall Trap
Learning to discriminate
Classical models inflated recall by calling almost everything positive. The QCNN kept 85% recall while actually raising precision — a meaningfully different behaviour.
03
Quantum Expressivity
High-dimensional feature space
Data Re-uploading and Projective Measurement expanded the output from 8 to 256 dimensions, enabling finer-grained discrimination between visually similar lesions.
06
Conclusion
A properly designed hybrid quantum-classical architecture — with Data Re-uploading and high-dimensional projective measurement — can outperform an equivalent classical model on tasks that demand generalization from limited data.
Not a clinical result, not a universal claim — a narrow, reproducible advantage on one real problem. Enough to justify pushing the architecture further.
Built with
PythonPyTorchPennyLaneResNet188-Qubit PQC
Related Research
Looking for the production model?
The high-performance ConvNeXtBase ensemble achieving 98.59% recall for clinical melanoma detection.
Read the
full work.
The repository holds every experiment, the failed iterations, the training scripts, and the notebooks used for the final comparison.
Type
Research project · MSc coursework
Architecture
ResNet18 · 8-Qubit PQC · Data Re-uploading
Stack
PyTorch · PennyLane