8.2.1. Neural networks

The development of ANNs stems from the anatomy and operations of the human brain. ANNs are made up of interconnected nodes or “neurons” that analyze input data, processing them through weighted connections (Zhang and Zhang 2018).

As a neural network undergoes more analysis, the hidden layers begin to extract increasingly complex hierarchical features, giving them the name deep neural networks (DNNs). DL algorithms obtain features that can be used for various classification, regression and recognition tasks belonging to different datasets.

Among the most popular include CNNs and RNNs, convolutional and recurrent neural networks, respectively. CNNs are primarily used for real-time image and video analysis, using convolutional layers for visualization of spatial hierarchies of visual content. In medical imaging, they have proven to be highly effective in oncology for tumor detection and organ segmentation (Alzubaidi et al. 2021). Other variations, such as ResNet and DenseNet, help improve the efficiency and accuracy of the learning task because of the special ways in which the nodes are connected.

Conversely, RNNs are specifically made for sequential or time-sensitive data. Due to their internal memory, they can keep context throughout sequences, which makes it easier for them to evaluate patient health records, bio-signals (ECG and EMG) and rehabilitation progress over time. More developed forms of RNNs, such as long short-term memory (LSTM) and gated recurrent units (GRU), alleviate problems associated with vanishing gradients and allow for learning of long-term dependencies (Mienye et al. 2024).

Furthermore, such novel designs such as transformers and graph neural networks (GNNs) are being used in healthcare due to their more effective work in natural language processing (NLP) and feature relational data modeling, respectively. These advancements increase the ability of neural networks to process complicated multimodal healthcare data, thus improving their usefulness in oncology and physiotherapy (Nogales et al. 2024).

8.2.2. Training and optimization

Training DL models comes with the need for large datasets, which require data to be fed into the model to adjust its parameters in order to minimize a loss function. This is defined mathematically as a quantified difference between anticipated results vis-a-vis the actual results (Menghani 2023).

The process of training a model involves forward propagation where operational data are run through the network until a prediction is reached. This is followed by backpropagation where the prediction is evaluated against reality and changes are made to the previously established weights via optimization algorithms.

Stochastic optimization methods, such as stochastic gradient descent (SGD), update the parameters using mini-batches of data to compute the gradient of the defined loss function (Mustapha et al. 2020). While SGD is relatively straightforward, it usually entails getting the learning rate to work and can be quite slow in terms of convergence. To solve these problems, ADAM (Adaptive Moment Estimation) has become more widely used. ADAM combines the benefits of two other SGD extensions, AdaGrad and root mean square propagation (RMSProp), by computing the first and second moments of the gradients and adjusting the learning rates at the per parameter level (Jie et al. 2022).

Other notable optimization techniques include NADAM (Nesterov-accelerated ADAM), AdaMax and AMSGrad, each providing trade-offs in terms of convergence speed and stability. Additionally, regularization methods such as dropout, L1/L2 metrics and batch normalization are incorporated during training to manage overfitting and ensure the model generalizes well (Thakkar and Lohiya 2021).

Furthermore, whether done manually or automatically through grid search, random search or Bayesian optimization, hyperparameter tuning becomes crucial in the context of complex healthcare applications tangentially aimed at achieving enhanced performance.

8.2.3. Evaluation metrics

Depending on the specific task, DL models are assessed through accuracy, precision, recall, F1 score and the area under the receiver operating characteristic curve (AUC-ROC) among others (Naidu et al. 2023). These metrics are tailored distinctly to the type, be it classification, regression, segmentation or ranking.

When classes are balanced, accuracy is useful and measures the percentage of correct predictions out of total predictions. Precision measures the fraction of actual positive results which when correctly predicted positive, reflect a greater accuracy to positive predictions (Powers 2020). In contrast, recall, also known as sensitivity, measures how well a model can recognize all relevant instances of a particular class, or the number of actual positives that were predicted to be positive. F1 score is another metric which is useful when the classes are imbalanced as it provides a balance along the spectrum of precision and recall.

AUC-ROC measures the ability of a model to tell the difference between the classes within all thresholds of classification. It is crucial for medical diagnostics when evaluating screening tests where both false positives and true test results have serious consequences (Varoquaux and Colliot 2023). For segmentation tasks, metrics such as the dice similarity coefficient (DSC) and intersection over union (IoU) are used to measure spatial overlap between the object’s predicted region and actual region.

Other metrics such as the Matthews correlation coefficient (MCC), Cohen’s kappa and even mean average precision (MAP) can be utilized for assessing a model’s performance at a deeper level. Brier score and calibration plots become extremely important in clinical settings where likelihoods derived from models influence decisions due to their classification model calibration value (Huang et al. 2020). In regression problems such as estimating recovery time, the discrepancy between the estimate and actual values is measured using mean squared error (MSE), root mean squared error (RMSE) and mean absolute error (MAE) metrics.

8.3.1. Early detection and diagnosis

The application of DL algorithms greatly enhances the chances of cancers being detected early, which improves prognosis and treatment outcomes significantly (Khandakar et al. 2024). For example, CNNs have been shown to be as accurate as trained radiologists at interpreting mammography images and recognizing breast cancer far earlier than it is clinically detected. These models detect and classify the type and grade of lesions that they can find.

Other than mammograms, DL has been applied to other imaging processes such as lung cancer screening using LDCT, brain cancer MRI and prostate cancer MRI (Chen et al. 2024). The automation of image interpretation is made easier by CNNs, and their more advanced counterparts such as VGGNet and ResNet possess a lot of sensitivity and specificity.

In addition, DL models are increasingly being used for interpreting histopathological images. With the development of whole slide imagers (WSIs), which hold a great deal of detail at the cellular level, DL algorithms can now efficiently identify cancerous cells, classify tumors and, in some cases, identify tumor-associated genetic alterations. This ability helps pathologists make reliable and timely diagnostic assessments (Erdur et al. 2025).

Recent innovations have also transformed cancer screening by incorporating DL with liquid biopsy and blood biomarker data, enabling noninvasive procedures. DL frameworks are advancing towards real-time multimodal comprehensive early detection systems that support clinical decision-making.

8.3.2. Tumor classification and segmentation

Effective treatment planning, surgical intervention and targeted radiation therapy greatly depend on the precise tumor classification and segmentation (Pranitha and Vurukonda 2024). The field of medical image segmentation has been driven by the adoption of DL methods, especially with the development of convolutional neural networks. For example, U-Net and its derivatives, such as Attention U-Net, 3D U-Net and nnU-Net. These architectures excel in localizing tumors in MRI, CT and PET scans, which require precise contouring of the region of interest.

In classification tasks, labeled datasets are used to train DL models for classifying the tumor within the image, such as identifying glioblastoma versus lower-grade gliomas in brain scans, or adenocarcinoma versus squamous cell carcinoma in lung images (Ali et al. 2021). Performance is further improved with hybrid models that integrate CNNs with support vector machines (SVMs) and with recurrent neural networks (RNNs) where spatial and temporal information are brought into the analysis.

Segmentation assists in visualizing the tumor’s boundaries, and it is important in radiomics where quantitative features derived from segmented areas are relevant to prognosis and treatment response. DL algorithms have been shown to segment the primary tumors, metastatic tumors and other surrounding anatomical structures, which is critical for protecting normal tissues during treatment.

In addition, the use of transfer learning and domain adaptation methods make it possible to customize the segmentation models to other departments or imaging approaches within different institutions, increasing the model’s applicability. Such developments highlight the promise of DL in increasing the accuracy and consistency of tumor classification and segmentation in clinical workflows within oncology (Hussain et al. 2024).

8.3.3. Triaging for treatment response prediction

The ability to predict patients’ response to treatment has been shown to greatly improve patient outcomes due to the possibility of customizing treatment or removing nonbeneficial interventions.

DL models are designed to evaluate a wide array of data attributes, from genetic and genomic data, through radiological imaging to structured electronic health records, in order to estimate the efficacy of treatment and its tolerance for the patient. For instance, models using pretreatment MRI or PET-CT scans have been shown to estimate chances for a favorable response to chemotherapy, radiation or immunotherapy, which helps clinicians to optimize the protocol selection (Selli et al. 2016).

Combining different biological omics disciplines such as transcriptomics, proteomics and metabolomics significantly improves the depth of predictive insight provided by DL models. These data provide critical insights into tumor biology and the patient-specific variability in drug metabolism. With regard to outcome prediction, attention and GNN-based DL models are capable of untangling many intricate connections within treatment results and molecular markers.

DL algorithms are also working to track treatment response over time by evaluating follow-up imaging and biomarker changes (Booth et al. 2022). This approach facilitates dynamic interventions, where changes to treatment plans occur in reaction to preliminary signs of resistance or considerable treatment efficacy. Predictive models have also been useful in elucidating some of the negative impacts of certain chemotherapeutic agents, such as cardiotoxicity, thus contributing to strategies aimed at reducing risks.

Related posts:

The Role of Artificial Intelligence in Personalized Physiotherapy and Cancer Treatment Model Validation Techniques for AI in Cancer Research Based on Physiotherapy and Oncology Machine Learning in Oncology and Physiotherapy: A New Era of Precision Medicine AI and Wearable Technology in Physiotherapy for Oncology Patients Smart Robotics in Physiotherapy and Oncology: Redefining Patient Outcomes Recurrent Neural Networks for Predictive Modeling in Cancer Time Series Data

Stay updated, free articles. Join our Telegram channel

Join