High-quality, low-quantity: A data-centric approach to deep learning performance optimization in digital X-Ray radiography
Bibliographic record
Abstract
The accuracy of identifying defects using specialized deep learning models can be affected by the circumstances in which the training data is curated. This is especially evident in digital X-ray radiography, where the depiction of flaws is significantly impacted by the exposure conditions. This study examines the effect of curating high-quality data on deep learning models. The variation in contrast-to-noise ratio (CNR), which is a crucial metric of image quality between features of interest and an adjacent normal background, has been found to be a key factor in model generalization in digital X-ray radiography applications. By making systematic alterations to exposure conditions during data curation, it was possible to obtain several representations of flaws in each test component with varying contrast-to-noise ratios (CNR) in the resultant radiographs. To evaluate the efficacy of the model under various conditions, two distinct datasets were curated. Dataset 1 was obtained by acquiring images with a consistent exposure setting on 140 test samples. The samples contained 4 morphologically distinct classes of flat bottom holes with seven different depths and sizes. The contrast-to-noise ratio (CNR) representations of flaws in this dataset can be attributed only to differences in depth in Dataset 1. Additionally, Dataset 2 was curated with an expanded range of CNR values by methodically adjusting exposure settings during image acquisitions. Hence, only 42 % of the test pieces from Dataset 1, which had three distinct depths of flat bottom holes, were used. Each of the two datasets was used to separately train YOLOv8 for instance segmentation and U-net for multi-class semantic segmentation. Each model was trained under the same conditions, and their performances were assessed using test sets from both dataset groups. The model trained on Dataset 1 exhibited a notable decline in performance when evaluated on test sets from Dataset 2, suggesting a lack of generalization ability. Conversely, the model that was trained using Dataset 2 consistently achieved high accuracy on both test sets, demonstrating impressive performance and successful generalization. This work shows that the generalization abilities of deep learning models may be improved by varying the contrast-to-noise ratio (CNR) of features in the training data. This finding paves the way for practical applications in digital X-ray radiography.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.000 | 0.000 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.000 |
| Open science | 0.000 | 0.000 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one teacher head, not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".