CheXVision

1Introduction

Deep convolutional networks trained on large annotated image collections have driven substantial gains in recognition, detection, and segmentation. Medical imaging has begun to benefit from the same paradigm, but it poses extra difficulties: pathological findings are often small and low-contrast relative to the full image, expert annotation is expensive, and ground-truth labels are frequently mined from free-text reports rather than verified by region-level markup.

Chest radiography is a natural testbed. It is the most common imaging exam in clinical practice, and a system that flags likely pathologies and separates normal from abnormal films would help radiologists prioritise their reading lists and extend basic screening to settings without specialist coverage. This motivation underlies the NIH ChestX-ray14 release (Wang et al., 2017) and the subsequent CheXNet result (Rajpurkar et al., 2017), which reported radiologist-level pneumonia detection and established DenseNet-121 as the reference architecture for this database.

The course brief requires two models on a shared dataset — one trained from scratch and one built with an alternative paradigm (transfer learning) — together with an explicit quantitative comparison. We frame the problem as two complementary supervised tasks on the same images. In the multi-label task the network predicts an independent probability for each of the fourteen pathologies; the labels are not mutually exclusive, so a single film may simultaneously show Effusion, Infiltration, and Atelectasis. This is the harder, clinically richer task and our primary evaluation target. In the binary task the network predicts a single Normal-versus-Abnormal score indicating whether any finding is present — the screening task, and a useful auxiliary signal on the shared features.

Both tasks are solved by one network with two heads on a shared backbone, so a single forward pass produces both predictions. This keeps the architecture comparison controlled: only the backbone changes, while task formulation, loss, data, and evaluation are held fixed. Throughout, we adopt the area under the ROC curve, macro-averaged over the fourteen pathologies, as the primary metric. AUC-ROC is threshold-independent and robust to severe class imbalance, and it is the measure reported by prior work on ChestX-ray14. Accuracy would be uninformative: a classifier that always predicts the absence of the rarest pathology, Hernia (prevalence well below 1%), scores over 99% accuracy on that class while detecting nothing of clinical value.

2Dataset and Big-Data Engineering

We use NIH ChestX-ray14 (Wang et al., 2017), one of the largest publicly available labelled chest X-ray collections: 112,120 frontal-view images from 30,805 unique patients, each tagged with zero or more of fourteen pathology labels mined from radiology reports by natural language processing. The labels are Atelectasis, Cardiomegaly, Effusion, Infiltration, Mass, Nodule, Pneumonia, Pneumothorax, Consolidation, Edema, Emphysema, Fibrosis, Pleural Thickening, and Hernia. Images with none of these findings are labelled "No Finding".

2.1Data partition

We shuffle the database once into a fixed 70/10/20 partition for training, validation, and testing, and store the assignment alongside the images so it is identical across every run and every model. The partition is frozen for reproducibility, and the test split is withheld from all model selection. Table 1 summarises the splits.

2.2Class imbalance

The dataset is severely imbalanced in two respects. First, roughly 54% of all images carry no finding, so the binary task is close to balanced (about 46% abnormal) while the multi-label task is dominated by negatives. Second, prevalence across the fourteen pathologies spans more than two orders of magnitude: Infiltration and Effusion are common, whereas Hernia appears in well under 1% of films. Left unaddressed, a network minimises its loss most easily by predicting "absent" for every rare class.

We counteract this directly with per-class positive weighting in the loss (Section 3.2). For each pathology we compute pos_weight = neg / pos on the training split and pass it to a numerically stable BCEWithLogitsLoss. For the rarest class the positive weight reaches roughly 500, so a single missed Hernia contributes about as much to the loss as 500 missed Effusions. This is the single most important choice for non-trivial performance on the rare tail.

2.3A pre-resized 320×320 dataset

The raw NIH release is about 45 GB of 1024×1024 PNGs — too large to stream efficiently from free cloud GPUs within a practical time budget. We therefore built a one-time preprocessing pipeline, run as a Kaggle kernel, that decodes the raw archives, resizes every image to 320×320 RGB, and writes 36 Parquet shards totalling about 7.97 GB, published as a versioned Hugging Face dataset (arudaev/chest-xray-14-320). The dataset revision hash is pinned in source, so every run reads byte-identical data. We keep 320×320 rather than the common 224-pixel baseline because chest pathology often hinges on small, low-contrast findings such as nodules and fine infiltrates; the higher resolution preserves that detail, at the cost of a smaller batch size to fit GPU memory.

3Methodology

3.1Dual-head, single-backbone design

Both models share an identical output design. A shared backbone extracts a feature vector once, and two small linear heads read from it: a multi-label head producing fourteen logits with an independent sigmoid per pathology, and a binary head producing a single Normal-versus-Abnormal logit. This choice is deliberate. Computing both tasks from shared features is cheaper than training two separate networks; the binary task acts as an auxiliary signal that regularises the shared representation; and, most importantly for the comparison, it isolates the variable under study. When we compare the from-scratch model against the DenseNet, only the backbone differs — heads, loss, data, and evaluation are held constant.

3.2Training procedure

Every component below is implemented in raw PyTorch, without a high-level training framework, so each decision is explicit and defensible. The settings are shared by both models except where Section 4.3 notes the two-phase schedule specific to the DenseNet.

3.3Data augmentation

Augmentation is chosen to reflect plausible radiographic variation rather than generic image noise; each transform maps to a real source of acquisition variation, as detailed below.

The code base also implements two optional, config-driven enhancements. CLAHE performs local contrast enhancement in LAB space to expose low-contrast findings, and label smoothing (ε=0.1) regularises against the noisy, report-mined labels. The final DenseNet checkpoint used both; the from-scratch checkpoint predates them — we flag this difference explicitly in Section 5.3, and closing the gap is the first item of future work.

4Models

4.1Model 1 — custom residual CNN, from scratch

The first model is a residual convolutional network built and trained entirely from scratch, with no pretrained weights of any kind. Its architecture follows ResNet-50-equivalent depth (He et al., 2016) and adds Squeeze-and-Excitation channel attention; it holds roughly 23 million parameters, all learned on chest X-rays alone.

The stem applies a 7×7 convolution at stride 2 followed by batch normalization, ReLU, and 3×3 max-pooling, capturing low-level edges and textures while reducing resolution fourfold. Four residual stages of depth [3, 4, 6, 3] follow. We chose ResNet-50 depth over a lighter [2,2,2,2] (ResNet-18) baseline because classifying fourteen simultaneous, often co-occurring signals needs more representational capacity, and the extra depth (about 23M against 11M parameters) is capacity worth its training cost. Each residual block is two 3×3 convolutions with a skip connection; the residual paths give gradients a short route and are what make a network this deep trainable at all, while batch normalization on every convolution stabilises training and permits a higher learning rate.

After each block, a Squeeze-and-Excitation module global-average-pools the feature map to a per-channel descriptor, learns per-channel importance through a small two-layer bottleneck with a sigmoid gate, and rescales the channels — channel-wise feature selection. A global average pool with dropout (0.5) then collapses the spatial dimensions without learned parameters, far less prone to overfitting than a large flatten-and-dense layer, and the dual heads (Linear 512→14 and Linear 512→1) read from the pooled vector. Convolutions use Kaiming (He) initialisation for the ReLU stack and the linear heads use Xavier; correct variance scaling is essential when training a deep network from random weights.

SE attention suits this problem well. Different pathologies activate different channels — the texture that signals Emphysema is not the one that signals Cardiomegaly. SE modules let the network amplify disease-relevant channels and suppress background channels per image, the kind of channel competition multi-label medical imaging creates, at negligible parameter cost (Hu et al., 2018). Training 23 million parameters from random initialisation on a noisy, imbalanced dataset is genuinely demanding — the network must learn edges, textures, and anatomy before disease, with no head start from natural-image pretraining. In return it provides a clean baseline against which the value of ImageNet transfer can be measured precisely (Section 5).

4.2Model 2 — DenseNet-121 transfer learning

The second model is a DenseNet-121 (Huang et al., 2017) pretrained on ImageNet and fine-tuned for our two tasks — the transfer-learning paradigm required by the brief, following the CheXNet design (Rajpurkar et al., 2017).

Three properties motivate this choice. Dense connectivity maximises feature reuse, strengthens gradient flow, and reaches strong accuracy with fewer parameters than a comparable ResNet — well suited to a free-GPU budget. ImageNet pretraining transfers well: low-level edges, gradients, and texture detectors are domain-agnostic, letting the network adapt to radiographic structure rather than relearn vision from scratch. Finally, the backbone is proven on this exact data: CheXNet reported a macro AUC-ROC of 0.841 with a single-head DenseNet-121, giving a literature-anchored target.

We adapt the backbone minimally for the dual-head design. The original 1000-class classifier is replaced by a shared projection layer (Linear 1024→512, ReLU, dropout 0.3) that maps the 1024-dimensional global-average-pooled output down to the 512-dimensional space shared with Model 1, followed by the same two task heads. This bridge layer is the only randomly initialised component; all backbone weights carry ImageNet features into training. The resulting trainable parameter count is approximately 7.9 million — the full backbone plus bridge and heads — against 23 million for the from-scratch model, yet it outperforms it by over four AUC points (Section 5).

4.3Two-phase fine-tuning

Fine-tuning a pretrained backbone with freshly initialised heads is delicate: the large early gradients from random heads can wash out, or catastrophically forget, the valuable ImageNet features. We avoid this with a two-phase schedule. In phase 1 (epochs 1–5) the backbone is frozen and only the feature layer and two heads are trained, at a learning rate of 1e-3, so the heads learn to read the fixed features without disturbing them. In phase 2 (from epoch 6) all layers are unfrozen at a tenfold lower 1e-4, letting the backbone gently specialise its mid- and high-level features to chest-X-ray textures without destroying what pretraining provided. This is the main reason the model converges quickly: its best validation checkpoint is at epoch 18, far earlier than the from-scratch model's epoch 41, because it starts from a strong initialisation rather than random noise.

5Experiments and Results

All numbers below are measured on the held-out validation split at each model's best checkpoint, selected by macro AUC-ROC. The multi-label task is evaluated with per-class and macro AUC-ROC, the binary task with AUC-ROC and F1.

5.1Overall comparison

Table 4 reports the headline metrics. The transfer-learned DenseNet is higher on every metric, reaching a macro AUC-ROC of 0.846 against 0.814 for the from-scratch network — a gain of 3.2 points. It also converges faster, reaching its best checkpoint at epoch 18 against epoch 41, despite fewer trainable parameters. The DenseNet edges past the CheXNet reference of 0.841, which we read as confirmation that our pipeline is sound rather than a meaningful improvement over that work.

5.2Per-class performance

Per-class AUC-ROC for the DenseNet (Table 5) reveals a clear, clinically sensible pattern: it is strongest on findings with a distinct global shape or high contrast, and weakest on diffuse, low-contrast ones.

The bottom of the ranking is occupied by Infiltration and Pneumonia, which are diffuse, subtle, and notoriously noisy in this dataset's report-mined labels. The low Pneumonia score in particular owes much to the scarcity of positives (under 1% of films); a lower ceiling on these classes is expected rather than a modelling failure, as human readers struggle with the same findings. Both models follow the same ordering — Edema, Cardiomegaly, and Emphysema easy; Infiltration and Pneumonia hard — which indicates the difficulty ranking is a property of the data (label noise, finding subtlety) rather than of either architecture. The models differ in ceiling: the DenseNet lifts the whole curve, most visibly on the mid-difficulty findings (Mass, Pneumothorax, Atelectasis), where richer pretrained features help separate subtle structure from background.

5.3Configuration of the reported runs

For full reproducibility we record the precise configuration of the two runs that produced the checkpoints above (Table 6), reported faithfully — including the difference in preprocessing and regularisation between them.

Both runs share identical preprocessing (CLAHE in LAB space, label smoothing ε=0.1), making this a fully controlled comparison. The 3.2-point AUC gap is attributable to the backbone and initialisation, not preprocessing differences.

5.4Inference-time enhancements

The evaluation and demo paths add two techniques at no training cost. Test-time augmentation averages the sigmoid probabilities over four views (original, horizontal flip, and ±7° rotations), reducing prediction variance. A two-model ensemble then averages the test-time-augmented probabilities of both models; because a from-scratch CNN and a pretrained DenseNet have different inductive biases and fail on different images, averaging their outputs is a standard, reliable way to recover a small amount of macro AUC over either alone.

6Discussion

What this project delivers is a controlled image-level classification comparison, not a clinical pathology detector. Holding data, task formulation, heads, loss, preprocessing, and evaluation fixed, the backbone and its initialisation are the sole key variables — which isolates the architecture and pretraining effect cleanly. Transfer learning outperforms from-scratch training by 3.2 macro AUC-ROC points while converging 2.3× faster and training fewer parameters. This is the expected outcome, reproduced cleanly on 112,120 medical images under a fully identical protocol. The from-scratch model is not a failure: at 0.814 macro AUC-ROC it has learned genuine pathology signal, and it grounds the transfer-learning gain in a concrete, controlled number.

6.1Limitations

• Classification, not localization. Both models output image-level probabilities only — no lesion bounding boxes or spatial maps. The demo is a screening classifier, not a detector.
• Moderate performance ceiling. The DenseNet reaches 0.846 macro AUC; binary AUC is 0.787 and F1 is 0.674. Informative, but not clinically actionable without further validation.
• Label noise. ChestX-ray14 labels are report-mined rather than annotator-verified, placing a hard ceiling on AUC — especially for diffuse findings such as Infiltration and Pneumonia.
• Validation figures only. All metrics are from the checkpoint-selection split; the 22,000-image test partition is held out and should be evaluated before any external claim.
• Single training run per model. No hyperparameter sweep, no cross-validation, and no repeat runs — stochasticity in optimisation means reported differences carry no confidence intervals.
• No external validation; educational use only. Trained on one institution's distribution; performance will degrade on different equipment or populations. Predictions must not substitute for professional medical judgement.

6.2Future work

• Compute final held-out test metrics with per-class confidence intervals.
• Repeat each run at least once to estimate variance and attach confidence intervals to the 1.7-point gap.
• Add Grad-CAM saliency overlays and quantify the ensemble's AUC gain; tune per-class decision thresholds for the screening task.
• Run longer experiments on university hardware and test higher-resolution input (448×448) for subtle findings such as nodules.
• Pursue lesion localisation if suitable annotated data becomes available — a distinct task from the current image-level classifier.

6.3Reproducibility and engineering

The cloud-first pipeline is a project strength: NIH source → Kaggle preprocessing → Hugging Face dataset → training kernels → model hub and Space demo. Runs are fully deterministic; hyperparameters live in versioned YAML; every push triggers ruff linting, 51 pytest unit tests, and mypy type checking; artefacts upload automatically to the hub.

7Conclusion

We compared a custom SE-ResNet trained from scratch (0.814 macro AUC-ROC) against a DenseNet-121 adapted by transfer learning (0.846) on the 112,120-image NIH ChestX-ray14 database. Both solve a fourteen-class multi-label task and a binary screening task through a shared backbone with two heads, under identical preprocessing. The transfer-learned model edges past the CheXNet reference and converges 2.3× faster; the from-scratch baseline grounds the gain in a concrete, controlled number. The system is reproducible, continuously tested, and deployed as a live demo. Fully automated classification from weakly supervised labels remains hard; §6.2 marks the natural next steps.

Acknowledgements

This work was carried out as project work for the Deep Learning and Big Data course in the AIN program. We thank Kaggle for the free GPU compute and Hugging Face for the dataset, model, and Space hosting that made the cloud-only workflow possible.

Appendix A · Individual Contributions

All four members contributed to discussion, design review, and the final report and presentation. Primary ownership of the work packages was distributed as follows.

This table reflects primary ownership; in practice the work packages overlapped and were reviewed collaboratively.

Appendix B · Project Resources

References

1. Wang, X., Peng, Y., Lu, L., Lu, Z., Bagheri, M., & Summers, R. M. (2017). ChestX-ray8: Hospital-scale Chest X-ray Database and Benchmarks on Weakly-Supervised Classification and Localization of Common Thorax Diseases. CVPR.
2. Rajpurkar, P., Irvin, J., Zhu, K., et al. (2017). CheXNet: Radiologist-Level Pneumonia Detection on Chest X-Rays with Deep Learning. arXiv:1711.05225.
3. Huang, G., Liu, Z., van der Maaten, L., & Weinberger, K. Q. (2017). Densely Connected Convolutional Networks. CVPR.
4. He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep Residual Learning for Image Recognition. CVPR.
5. Hu, J., Shen, L., & Sun, G. (2018). Squeeze-and-Excitation Networks. CVPR.
6. Loshchilov, I., & Hutter, F. (2019). Decoupled Weight Decay Regularization. ICLR.