Delta Radiomics for Tracking Dynamic Tissue Changes

43 concepts32 papers1 proposal

Traditional radiomics extracts hundreds of quantitative features—such as texture, shape, and intensity—from a single medical image. While powerful, this static approach treats a tumor or tissue structure as a frozen snapshot in time. In reality, biology is fundamentally dynamic: tumors grow, blood flows, contrast agents wash in and out, and tissues mutate or shrink under the pressure of therapy.

Delta radiomics (and its counterpart, temporal radiomics) shifts the analytical paradigm from static snapshots to the trajectory of feature changes across multiple scans. By computing the mathematical variance or rate of change between imaging states, models can isolate the underlying biological response from background anatomical noise.


Longitudinal Delta Radiomics: Tracking Treatment Trajectories

When a patient undergoes therapy, a baseline scan only reveals where the disease started, not how it will behave. By analyzing the longitudinal evolution of features before, during, and after treatment, machine learning models gain a direct window into tumor kinetics and radiosensitivity.

[Baseline Scan] ------(Therapy / Time)------> [Post-Treatment Scan]
       |                                              |
[Static Features]                             [Static Features]
       \                                              /
        \---> [Mathematical Delta / Trajectory] <----/
                             |
              (Isolates Biological Response)

This longitudinal tracking captures the dynamic evolution of tissue characteristics, which provides a superior predictive signal for long-term survival and treatment efficacy compared to static baseline assessments. For instance, Jiahui Wang et al. demonstrated that tracking feature shifts over serial PET/CT scans allows models to distinguish between treatment-induced inflammation and persistent disease—a classic clinical conundrum that static scans fail to resolve.

Similarly, in predicting responses to neoadjuvant chemotherapy or immunotherapy, observing the initial trajectory of a tumor under therapy captures biological sensitivity that baseline snapshots miss (Jianming Guo et al.; Sunyi Zheng et al.). To quantify these shifts, researchers often calculate the relative net change in features (Ming Fan et al.) or integrate tumor growth kinetics, such as the log volume change rate, directly into deep learning architectures (Guangjie Liu).

This dynamic tracking has proven highly effective across various oncological applications, including:

  • Pediatric Gliomas: Predicting survival by comparing diagnostic scans with post-radiation changes (Xinyang Liu et al.).
  • Cervical Cancer: Assessing lymph node metastasis and disease-free survival by quantifying tumor dynamics and radiosensitivity during therapy (Jinjin Liu et al.; Weiping Wang et al.).
  • Colorectal Cancer: Evaluating long-term patient outcomes through the temporal evolution of biological structures (Xiangyu Liu et al.).

Ultimately, these insights support adaptive therapy, allowing clinicians to modify treatment plans in real-time based on the evolving state of the tumor (Arnaud Beddok et al.).


Multi-Phase Radiomics: Capturing Hemodynamic Flux

The temporal dimension is equally valuable on a shorter timescale: the seconds to minutes it takes for a contrast agent to circulate through the body. Relying on a single contrast phase, such as the venous phase, can exclude critical diagnostic information regarding tumor vascularity and hemodynamics (Yating Wang et al.).

Multi-phase temporal radiomics extracts features across non-enhanced, arterial, venous, and delayed phases. This captures the complex wash-in and wash-out dynamics of contrast agents, which directly reflect physiological properties like tissue density, blood flow, and capillary permeability (Qingbo Yin et al.; Li Gao et al.).

[Contrast Injection] ---> [Arterial Phase] ---> [Venous Phase] ---> [Delayed Phase]
                                 |                     |                  |
                                 +----------+----------+------------------+
                                            |
                             [Multi-Phase Feature Fusion]
                                            |
                             (Maps Hemodynamic Signature)

By fusing data across these temporal phases, models map out the tumor microenvironment and predict biological markers of proliferation like Ki-67 expression (Tianjun Song et al.) or profile the tumor immune landscape (the authors of A predictive radiotranscriptomics model based on DCE-MRI for tumor immune landscape and immunotherapy in cholangiocarcinoma).

This multi-temporal approach has been successfully applied to improve diagnostic and stratification accuracy across multiple domains:

  • Pancreatic and Liver Lesions: Stratifying pancreatic cystic neoplasms (Xiaohan Yuan et al.) and characterizing hepatocellular carcinoma on contrast-enhanced ultrasound (Xiaoxin Liang et al.).
  • Thoracic and Breast Imaging: Differentiating lung nodules (Na Liu et al.) and automating breast cancer diagnosis using dynamic contrast-enhanced sequences (Guangwu Qian et al.).
  • Renal Carcinomas: Evaluating the tissue matrix and collagen fiber content based on how contrast agents are retained or cleared (Zhongyuan Li et al.).

Furthermore, in dynamic PET imaging, calculating voxel-wise parametric maps of tracer movement (such as time-to-peak and uptake slope) allows models to distinguish molecular subtypes based on metabolic flux rather than static uptake alone (Lena Kaiser et al.).


Why It Works: The Power of Differential Filtering

Why does calculating the change between states consistently outperform a single high-resolution snapshot? The answer lies in differential filtering.

Every patient's anatomy contains unique background noise—variations in baseline tissue density, benign structural quirks, or patient-specific metabolic rates. When a machine learning model looks at a single scan, it must attempt to separate this complex anatomical noise from the actual pathological signal.

Delta radiomics bypasses this problem by using the patient as their own control. For example, Louise Nissen demonstrated that subtracting quantitative features extracted from a baseline rest state from those extracted during a physiological stress state isolates the specific tissue response to a stimulus.

Mathematically, if a scan $S$ is composed of static anatomy $A$ and a dynamic biological process $P$:

$$S_{\text{state 1}} = A + P_{\text{state 1}}$$

$$S_{\text{state 2}} = A + P_{\text{state 2}}$$

By subtracting the two states to find the delta ($\Delta$), the static anatomy $A$ cancels out:

$$\Delta = S_{\text{state 2}} - S_{\text{state 1}} = P_{\text{state 2}} - P_{\text{state 1}}$$

This subtraction filters out the static anatomical noise, leaving a clean signal that represents pure functional, hemodynamic, or therapeutic change. This trajectory-based signal provides a far more robust and generalizable foundation for machine learning models than static snapshots.

Key Foundational Work

MRI time series modeling of MS lesion development (2006) — contributed 1 concept elements to this direction.
Ultrasound-Based Characterization of Prostate Cancer: An in vivo Clinical Feasibility Study (2013) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_delta_radiomics_efe54490,
  title = {delta radiomics},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/delta-radiomics/rdr_efe5449052f043cd8a6d2db24880fad0},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_delta_radiomics_efe54490_v1,
  title = {delta radiomics},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/delta-radiomics/rdr_efe5449052f043cd8a6d2db24880fad0/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake