Concept · artificial-intelligence
Convolutional neural networks
5 studies1 discoveryEvidence last moved Sep 27, 2026
A convolutional neural network learns small filters that slide across an input (an image, a 3D grid or a string of characters) to detect local patterns, then stacks them into higher-level features. The evidence here is from computational studies applying CNNs to echocardiograms, histopathology slides, chemical SMILES strings, cancer drug response and 3D protein structures.
CNNs often post impressive headline numbers in biomedicine. These studies show both where that performance is real and how data splits and leakage can inflate it.
Studies
5
Findings
5
5 supporting · 0 challenging · 1 qualifying citations
Open tensions
1
Latest change
Concept page published
Convolutional neural networks
Currently
What we know
- CNNs can learn genuine image structure for medical view recognition.
- Transferred CNN features can outperform hand-engineered ones even without much retraining.
- CNNs can learn chemical motifs from text, but expert features still competed.
- Random splits can measure table-filling, not prediction for new drugs.
- Similar sequences in train and test sets make CNNs look better than they are.
Largest unresolved question
Image tasks with clear visual structure (echocardiogram views, tumour histology) produced very high accuracies, while molecular and protein tasks gave moderate or split-dependent results; these are different data types and evaluation setups, not a direct conflict.
Common misconceptions
If a CNN takes chemical structure as input, it must be using chemistry to predict.
In the drug-response study, scrambling SMILES barely changed accuracy and performance collapsed on unseen drugs, so the network was exploiting statistical patterns rather than chemical meaning.
A random train/test split is always a fair test.
Random splits let the same drugs, cell lines or homologous proteins appear on both sides; held-out drugs and homology pruning gave much lower scores.
Related
Claim ledger
What the evidence shows
Drawn from 5 studies in this library. Mix labels say which citation roles are present; they are not a strength score. Supports means evidence for a finding; Challenges means evidence against a stated position; Qualifies marks scope.
CNNs can learn genuine image structure for medical view recognition.
On clinical echocardiograms from one US centre, a CNN classified 15 views with 97.8% accuracy on videos and 91.7% on single stills, versus 79.4% for four experts on the same low-resolution stills; with shuffled labels it fell to 6.9%.
- Can a neural network recognise which heart-ultrasound view it's seeing?— Single centre, single labeller, no external validation; experts saw an unfamiliar image format.
Transferred CNN features can outperform hand-engineered ones even without much retraining.
Features from an ImageNet-pretrained CNN beat hand-crafted features by about 20% in brain-tumour histopathology classification and won the MICCAI 2014 challenge at 97.5% accuracy, while fine-tuning on histology added only small gains.
CNNs can learn chemical motifs from text, but expert features still competed.
A CNN on SMILES strings beat logistic regression, random forest and fingerprint-based networks in Tox21 cross-validation, and one filter picked out a steroid-like motif, but it fell slightly short of models that added thousands of hand-crafted toxicophore features.
Random splits can measure table-filling, not prediction for new drugs.
A CNN predicting cancer drug response reached Pearson 0.909 on random drug-cell-line pair splits but about 0.2 when whole drugs were held out, and shuffling SMILES barely changed accuracy.
Similar sequences in train and test sets make CNNs look better than they are.
A 3D CNN for protein stability changes reached correlations near 0.47, and a version trained without removing homologous proteins scored higher on the same test (reverse r 0.59 vs 0.48), showing leakage inflates results.
Debates
Tensions and limits
Some items are genuine disagreements on the same question. Others mark different assays, populations, or outcomes.
Image tasks with clear visual structure (echocardiogram views, tumour histology) produced very high accuracies, while molecular and protein tasks gave moderate or split-dependent results; these are different data types and evaluation setups, not a direct conflict.
Image tasks with clear visual structure (echocardiogram views, tumour histology) produced very high accuracies, while molecular and protein tasks gave moderate or split-dependent results; these are different data types and evaluation setups, not a direct conflict.
- Can a neural network recognise which heart-ultrasound view it's seeing?
- Can ImageNet-trained networks read cancer tissue slides?
- Can a CNN predict how cancer cells respond to drugs?
- Can a 3D image network predict how mutations change protein stability?
Study Role Design N Population Outcome Can a neural network recognise which heart-ultrasound view it's seeing? Supports Computational / modellingSupervised CNN trained on 15-view labelled echocardiogram frames, with train/validation/test split by patient study, a random-label negative control, and a comparison with 4 human experts. N=267 · Echocardiogram studies from different patients: 240 for training and validation (213 train, 27 validation) plus 27 for testing, yielding 223,787 labelled frames; 4 board-certified echocardiographers each classified 1500 test images. Randomly selected clinical transthoracic echocardiograms from UCSF, 2000-2017, adults aged 20-96 Per-view and overall classification accuracy, F-score and AUC on held-out images and videos; comparison with expert accuracy Can ImageNet-trained networks read cancer tissue slides? Supports Computational / modellingAlexNet features from ImageNet extracted on image patches, pooled per slide and classified with a linear SVM; compared with hand-crafted features, whole-image CNN input, fine-tuning and prior methods Several datasets: a brain-tumour challenge set (22 low-grade and 23 glioblastoma training slides for classification, 35 for segmentation) and a colon-cancer set of 717 cropped regions; no single N H&E-stained histopathology images of brain tumours (glioblastoma vs low-grade glioma) and colon cancer subtypes Classification accuracy and overlap-based segmentation accuracy Can a CNN predict how cancer cells respond to drugs? Supports Computational / modellingRegression of normalised IC50 from one-hot SMILES and 735 genetic features; random 80/10/10 pair splits repeated 50 times, plus drug-blind, cell-line-blind and tissue-blind splits. N=172114 · 172,114 measured drug-cell line IC50 values among 223 drugs and 948 cell lines from GDSC. Anti-cancer drugs and cancer cell lines in the Genomics of Drug Sensitivity in Cancer database Coefficient of determination, Pearson correlation and RMSE between predicted and measured IC50 Can a 3D image network predict how mutations change protein stability? Supports Computational / modellingEnsemble of ten 3D CNNs on voxelised wild-type and mutant structures, trained on homology-pruned data augmented with reverse mutations; compared with seventeen predictors on a balanced blind test set, plus ablations and case studies. N=684 · 684 test mutations in S_sym (342 direct plus their reverses); training used 3,488 points (1,744 direct mutations plus reverses); case studies of 42 p53 and 134 myoglobin mutations; ClinVar variants for an application analysis. Experimentally measured single-point protein mutations and their structures Pearson correlation and RMSE between predicted and measured stability change (ΔΔG), and anti-symmetry bias between direct and reverse mutations
PaperFren reads this as a limit on how far one study travels — different assays, populations, or outcomes — not a forced fight between papers.
Timeline
How understanding moved
Study years are when the paper was published. Evidence edits are dated changes to this page's claims. Explanations are when PaperFren added a Discovery — not a claim that the science happened that day.
2026
- Deep networks trained on object labels predict visual cortex better than hand-built and unsupervised models
Concept page published
Convolutional neural networks
Change log
What changed
Dated edits to this page's evidence: studies added or removed from a claim, claims added or withdrawn, and new explanations tagged here. Rewordings are not listed.
- Concept page published
Papers
5 studies in this library bear on Convolutional neural networks, ordered by citations.
- Can a neural network recognise which heart-ultrasound view it's seeing?
A convolutional neural network named 15 standard echocardiogram views more accurately than board-certified experts, using tiny low-resolution images.
- Can ImageNet-trained networks read cancer tissue slides?
Features borrowed from a network trained on everyday photos, applied to small patches of huge tissue slides, classified and outlined cancer far better than hand-designed image features.
- Can a 3D image network predict how mutations change protein stability?
Training a 3D convolutional network on both each mutation and its reverse removed the usual bias toward predicting destabilisation, and removing related proteins from training showed earlier methods' scores were likely inflated.
- Can a CNN predict how cancer cells respond to drugs?
A two-branch convolutional network predicted drug response well for drugs it had seen before, but almost failed on entirely new drugs.
- Can a text-style CNN read chemical formulas to predict toxicity?
Treating a molecule's SMILES text string like a one-dimensional image let a CNN beat standard fingerprint methods on toxicity prediction and point to the substructures it relied on.
Compare studies
Select 2–10 studies. Design and N are labels, not a ranking.
Nothing selected yet.
Questions
What is still open
Image tasks with clear visual structure (echocardiogram views, tumour histology) produced very high accuracies, while molecular and protein tasks gave moderate or split-dependent results; these are different data types and evaluation setups, not a direct conflict.
Ask PaperFren about Convolutional neural networks
Study this conceptflashcards and short-answer questions
Describe two sanity checks that show a CNN learned something real, citing studies.
A shuffled-label control: the echocardiogram CNN dropped to 6.9% accuracy, near chance for 15 classes, when labels were randomised. Occlusion: masking the heart, but not other regions, sharply reduced accuracy. Conversely, in drug-response prediction, scrambling SMILES barely changed accuracy, a warning that the model was not reading chemistry.
How can data leakage inflate CNN performance in protein studies?
If proteins similar in sequence appear in both training and test sets, the network can memorise rather than generalise. ThermoNet scored higher (reverse r 0.59 vs 0.48) when trained without homology pruning. After pruning, two other methods outperformed it, so fair splits change the ranking too.