Science News & Breakthroughs May 28, 2026

Artificial Intelligence in Cancer Detection and Diagnosis: Deep Learning Approaches in Modern Oncology

The mathematics behind deep learning in oncology, what trials and retrospective studies show, and why validation, calibration and dataset shift decide clinical value.

SC
ScienceTrace Science Desk
•
 8 min read
•
 1,597 words

Cancer is a leading cause of death worldwide, according to the World Health Organization, and for many cancers the stage at which the disease is found strongly influences outcome. That is the clinical motivation for applying deep learning to the images and records produced during screening and diagnosis. The mathematics behind these models is compact, and the technical results can be impressive. The harder questions are about whether a model that performs well on a research dataset helps patients when it is used in a hospital. This article walks through the core equations, then examines the evidence and its limits.

From weighted sums to a classifier

The basic unit of a neural network computes a weighted sum of its inputs, adds a bias and applies a non-linear activation function:

\[y=f\left(\sum_{i=1}^{n} w_i x_i+b\right)\]

Here:

  • \(x_i\) = input features
  • \(w_i\) = weights
  • \(b\) = bias
  • \(f\) = activation function
  • \(y\) = predicted output

A deep network stacks many layers of such units, and training adjusts the weights so that the output matches known labels, such as "malignant" or "benign" as determined by pathology. The network is only as good as those labels. If the reference standard is itself uncertain, for example when radiologists disagree about a subtle finding, the model inherits that noise.

Convolution: how image models see tissue

Medical images are large grids of pixel values, so networks that treat each pixel as an independent input would be unwieldy. Convolutional neural networks (CNNs) instead slide small filters across the image, so the same pattern detector is applied everywhere. The operation is usually written as:

\[S(i,j)=(I*K)(i,j)=\sum_m \sum_n I(m,n)K(i-m,j-n)\]

Here:

  • \(I\) = input medical image
  • \(K\) = convolution kernel
  • \(S(i,j)\) = feature map output

A technical aside: most deep learning libraries actually implement cross-correlation, which is the same sum without flipping the kernel. Since the kernel values are learned, the distinction does not matter in practice. Early layers learn edges and textures, deeper layers combine them into shapes and patterns, and a final layer produces a prediction. The 2017 study by Esteva and colleagues on skin lesions is a well-known example: a CNN trained on a large set of clinical images classified skin lesions with performance comparable to that of tested dermatologists on selected tasks. The claim is narrower than it sounds, as the comparison used curated test images, not the full workflow of a clinic.

Application areas where deep learning has been studied include:

  • Breast cancer screening on mammograms
  • Lung nodule detection and malignancy estimation on CT scans
  • Skin lesion classification from photographs and dermoscopy
  • Brain tumour segmentation on MRI, where the BraTS benchmark challenges have provided shared data for comparison
  • Pathology, where models analyse gigapixel whole-slide images split into small tiles

Probabilities and the sigmoid

For a yes-or-no question, the final output is typically passed through the sigmoid function, which maps any real number to a value between 0 and 1:

\[\sigma(z)=\frac{1}{1+e^{-z}}\]

Here:

  • \(z\) = weighted input value
  • \(\sigma(z)\) = probability prediction

The output behaves like a probability only if the model is well calibrated, meaning that among cases scored at 0.8, about 80 percent are truly positive. Deep networks are often overconfident, so calibration should be measured separately from accuracy. A threshold converts the score into a decision, and where to set it is a clinical choice, not a mathematical one, because it trades missed cancers against false alarms.

Loss: what the model is optimised for

Training minimises a loss function. For binary classification the standard choice is binary cross-entropy:

\[L=-\frac{1}{N}\sum_{i=1}^{N}\left[y_i\log(\hat{y}_i)+(1-y_i)\log(1-\hat{y}_i)\right]\]

Here:

  • \(y_i\) = actual label
  • \(\hat{y}_i\) = predicted value
  • \(N\) = number of samples

A lower loss means predictions are closer to labels on the data being measured. It does not by itself mean the model will detect cancer reliably on new patients. Loss on the training data can fall almost to zero while performance on unseen data stays poor, which is why evaluation must use data that were held out, ideally from different hospitals and scanners.

Why a good score does not settle the question

Consider screening, where cancer is uncommon. As a purely illustrative calculation, not taken from any study, suppose 0.5 percent of people screened have the disease, and a model has 90 percent sensitivity and 95 percent specificity. Of 10,000 people, 50 have cancer and the model flags 45 of them. Of the 9,950 without cancer, 5 percent, about 498, are flagged falsely. So only about 45 of roughly 543 flagged cases, around 8 percent, are true cancers. The numbers that matter to patients, the positive predictive value and the downstream cost of false alarms, depend on prevalence as much as on the model.

What the clinical evidence shows

Several results are worth knowing. McKinney and colleagues (2020) evaluated an AI system on mammograms from the UK and the US and reported that it performed at least as well as the radiologists in their comparison, with reductions in both false positives and false negatives in the retrospective data, while noting that results varied across datasets. Ardila and colleagues (2019) described a deep learning model for lung cancer screening on low-dose CT that matched or exceeded radiologists in their evaluation. In pathology, Campanella and colleagues (2019) showed that models trained on slide-level labels could reach clinical-grade performance in prostate cancer detection on a large dataset. The MASAI randomised trial in Sweden (Lång and colleagues, 2023) is notable because it tested AI-supported mammography reading prospectively in a screening programme. It found that cancer detection was comparable to standard double reading, with a substantial reduction in radiologist screen-reading workload. Randomised, prospective evidence of this kind remains uncommon compared with the large number of retrospective studies.

Personalised treatment: more speculative

Models that combine genomic mutations, tumour features and treatment history to predict drug response or survival are an active area, but the evidence is less mature than for imaging. Small cohorts, batch effects between sequencing centres and survivorship bias in available records limit what can be inferred. A prediction of survival is also not a treatment recommendation. Showing that a model chooses a better therapy requires comparing outcomes in a prospective study.

Limitations and open questions

  • Dataset shift. Zech and colleagues (2018) found that a pneumonia detection model's performance on chest radiographs varied across hospitals, partly because the model learned site-specific cues. Equivalent problems occur in oncology imaging when scanner type or protocol changes.
  • Shortcut learning. Models can latch onto artefacts instead of pathology. Winkler and colleagues (2019) reported that surgical skin markings on lesions increased the false-positive rate of a dermoscopy CNN. Geirhos and colleagues (2020) review the general phenomenon.
  • Representation. Training sets that under-represent certain skin tones, ethnic groups, ages or imaging devices may lead to uneven performance. Several skin image datasets have historically contained few images of darker skin.
  • Interpretability. Heat maps indicating where a model "looked" can be misleading. Whether they help clinicians is an empirical question.
  • Overdiagnosis. A system that finds more small lesions may detect cancers that would never have caused harm, which can lead to unnecessary treatment.
  • Privacy and data access. Sharing medical data across institutions is restricted, which has encouraged work on federated learning, but that approach has its own technical and governance hurdles.

What researchers should watch

  • Prospective and randomised evaluations that report patient outcomes such as interval cancers and stage at diagnosis, not just model accuracy.
  • External validation across scanners, hospitals and countries, with subgroup reporting.
  • Reporting standards such as the TRIPOD and CONSORT-AI extensions that make studies comparable.
  • How radiologists and pathologists behave when a model's suggestion is shown: automation bias, in which clinicians over-trust a tool, can cancel out technical gains.
  • Performance in low-resource settings, including South Asia, where screening infrastructure and imaging equipment differ from those in the datasets used in most studies.

Conclusion

Deep learning has turned several diagnostic tasks into well-posed pattern recognition problems, and the best studies suggest it can match specialists on defined tasks or reduce their workload. The remaining gap is between benchmark and bedside. Closing it takes careful validation, honest reporting of errors and designs in which a clinician stays accountable for the decision. This article is general information for researchers and is not medical advice.

References and further reading

  1. World Health Organization. Cancer Fact Sheet. https://www.who.int/news-room/fact-sheets/detail/cancer
  2. LeCun, Y., Bengio, Y., Hinton, G. "Deep learning." Nature, 521, 436–444, 2015.
  3. Esteva, A., et al. "Dermatologist-level classification of skin cancer with deep neural networks." Nature, 542, 115–118, 2017.
  4. McKinney, S. M., et al. "International evaluation of an AI system for breast cancer screening." Nature, 577, 89–94, 2020.
  5. Ardila, D., et al. "End-to-end lung cancer screening with three-dimensional deep learning on low-dose chest computed tomography." Nature Medicine, 25, 954–961, 2019.
  6. Campanella, G., et al. "Clinical-grade computational pathology using weakly supervised deep learning on whole slide images." Nature Medicine, 25, 1301–1309, 2019.
  7. Lång, K., et al. "Artificial intelligence-supported screen reading versus standard double reading in the Mammography Screening with Artificial Intelligence trial (MASAI)." The Lancet Oncology, 24, 936–944, 2023.
  8. Zech, J. R., et al. "Variable generalization performance of a deep learning model to detect pneumonia in chest radiographs." PLOS Medicine, 15(11), e1002683, 2018.
  9. Winkler, J. K., et al. "Association between surgical skin markings in dermoscopic images and diagnostic performance of a deep learning convolutional neural network for melanoma recognition." JAMA Dermatology, 155, 1135–1141, 2019.
  10. Geirhos, R., et al. "Shortcut learning in deep neural networks." Nature Machine Intelligence, 2, 665–673, 2020.
  11. Menze, B. H., et al. "The Multimodal Brain Tumor Image Segmentation Benchmark (BRATS)." IEEE Transactions on Medical Imaging, 34, 1993–2024, 2015.

ScienceTrace — AI • Oncology • Scientific Research • Future Technology

#AI #cancer detection #deep learning #oncology #machine learning #CNN #medical imaging
All Scientific Breakthroughs