Triple-negative breast cancer (TNBC) is the most aggressive breast cancer subtype, defined by the absence of estrogen receptor, progesterone receptor, and HER2 expression. Because it lacks all three of the most common therapeutic targets, it cannot be treated with hormone therapy or HER2-directed drugs. TNBC has the worst prognosis of all breast cancer subtypes and the highest rate of early relapse and metastasis.
Diagnosis currently relies on tissue biopsy followed by immunohistochemistry (IHC) staining to detect receptor expression -- an invasive, time-consuming procedure. Most TNBC tumors also lack the typical radiological features visible on mammograms, making imaging-based detection unreliable. There is a growing need for novel, non-invasive approaches that can identify TNBC at the cellular level without requiring tissue labels or chemical stains.
Dynamic light scattering (DLS) offers a fundamentally different approach. When laser light illuminates a cell, it interacts with the cell's internal structures -- organelles, the cytoskeleton, the nucleus -- and scatters in patterns that fluctuate over time as these structures move. Because different cell types have different internal activity levels, the way scattered light intensity changes with time encodes information about what type of cell is being observed. DLS measures this temporal fluctuation pattern non-invasively and without any labels or dyes.
A key insight driving this study is the role of mitochondria in cancer cell behavior. Mitochondria -- the organelles that generate cellular energy -- are known to undergo rapid cycles of splitting (fission) and merging (fusion) in cancer cells. In TNBC specifically, enhanced mitochondrial fission generates smaller, more numerous mitochondria that accelerate cell proliferation, inhibit programmed cell death (apoptosis), and promote invasion and metastasis. This heightened mitochondrial activity produces measurable differences in light scattering behavior that, with the right analytical tools, can distinguish TNBC cells from other cancer subtypes.
The study built a custom dynamic light scattering microscopy system using a 532 nm laser focused through an objective lens and coupled into a single-mode optical fiber, which delivered the light to a microfluidic chip containing the cell suspension. A camera captured scattered light at 103.09 frames per second for approximately 10 seconds per cell, generating a short video of roughly 1,000 time-series image frames. The camera's objective lens was deliberately placed in a defocusing mode -- slightly out of focus from the cells -- so that scattered light from inside the cell spreads into a speckle pattern on the camera sensor. As subcellular structures move, these speckle patterns fluctuate, encoding the internal dynamics of the cell in the video signal.
Two breast cancer cell lines were used: MDA-MB-231 (TNBC, highly metastatic) and SKBR3 (HER2-positive, moderate metastatic potential). These represent two biologically distinct subtypes with different mitochondrial behavior. A total of 429 MDA-MB-231 videos and 372 SKBR3 videos were collected for the classification experiments.
To automatically classify the cell type from DLS videos, a two-stream deep learning framework was developed. This architecture was adapted from video action recognition research and processes two different types of information simultaneously. The first stream -- the spatial stream -- takes the first frame of each video and extracts visual features describing the cell's appearance: size, shape, texture, and density. The second stream -- the temporal stream -- processes optical flow images derived from consecutive video frames. Optical flow maps the apparent motion of patterns between frames, directly representing how the speckle patterns inside the cell are shifting over time.
Both streams use ResNet-50 as the backbone -- the same 50-layer residual network architecture used in prior medical imaging studies, fine-tuned here from ImageNet pre-trained weights. The spatial and temporal streams are trained separately and their outputs combined for final classification. This dual-stream design allows the model to leverage both static cell morphology (spatial information) and dynamic subcellular motion (temporal information) simultaneously, capturing more information than either stream alone.
The study first verified that DLS microscopy can detect intracellular motion at all. Live TNBC cells (MDA-MB-231) were compared with cells that had been chemically fixed with paraformaldehyde -- a preservative that arrests all cellular activity. The autocorrelation function of the scattered light signal -- a mathematical measure of how long the light pattern stays similar to itself over time -- decayed much more slowly for fixed cells than for live cells. This confirmed that the fluctuations in live cells reflect genuine intracellular movement, not imaging noise.
To directly test the role of mitochondria, the study used mdivi-1, a small molecule that inhibits the protein Drp1, which is responsible for driving mitochondrial fission. When TNBC cells were treated with mdivi-1, their mitochondria were prevented from dividing into smaller fragments. The autocorrelation function of the DLS signal from these inhibited cells decayed more slowly (decay rate 1.144 per second) than untreated TNBC cells (decay rate 1.462 per second). This quantitative shift confirms that mitochondrial fission activity directly contributes to the DLS signal -- slower fission means slower subcellular motion means a slower-decaying autocorrelation function.
Confocal fluorescent microscopy directly visualized the mitochondria in both cell lines using a red fluorescent dye (Mito-Tracker Red). The images showed that TNBC cells had fragmented mitochondria -- many small, dispersed mitochondria scattered throughout the cell -- while HER2-positive SKBR3 cells had interconnected mitochondrial networks -- elongated, connected structures spanning larger regions. A structural similarity analysis confirmed that mitochondrial patterns changed faster in TNBC cells over 20-second intervals, quantitatively confirming greater mitochondrial dynamics.
The DLS autocorrelation analysis of the two cell lines showed a consistently higher decay rate for TNBC cells (1.462 per second) versus HER2-positive cells (1.195 per second). A higher decay rate means faster scattering fluctuations, reflecting the more active, fragmented mitochondrial network in TNBC. This measurable difference in light scattering dynamics -- caused by biological differences in cancer aggressiveness -- forms the physical basis for the cell classification approach.
The two-stream deep learning framework was evaluated using five-fold cross-validation on 801 cell videos (429 TNBC, 372 HER2-positive). The model achieved an AUC of 0.95, accuracy of 0.89, sensitivity of 0.91, and specificity of 0.88 for distinguishing TNBC from HER2-positive cells.
The temporal stream's contribution was validated by comparing the two-stream model against a static light scattering baseline that used only a single image per cell (equivalent to the spatial stream alone). The static baseline achieved an AUC of only 0.75 and accuracy of 0.66 -- substantially lower than the dynamic approach. This comparison isolates the value of motion information: the temporal (optical flow) component contributed the majority of the model's discriminative power, increasing AUC by 0.20 and accuracy by 23 percentage points.
Optical flow images made the difference visually apparent. When the horizontal and vertical components of pixel movement were mapped between consecutive video frames, TNBC cells and HER2-positive cells showed clearly distinct patterns. TNBC cells had more dramatic, widespread intracellular motion, while HER2-positive cells showed less movement -- consistent with their slower, more network-like mitochondrial dynamics. These qualitative differences are precisely what the temporal stream quantifies and feeds to the classifier.
Optical flow is a computer vision technique that estimates how patterns of light intensity in an image appear to move from one frame to the next. For each pixel in the image, it computes a vector describing where that pixel's pattern seems to have shifted between the current and next frame. The result is a dense map of motion across the entire image -- showing both the direction and magnitude of apparent movement everywhere simultaneously.
In the context of DLS cell videos, the speckle patterns inside a cell represent subcellular structures scattering light. As mitochondria, vesicles, and other organelles move, they shift the speckle pattern. The optical flow of consecutive DLS frames therefore maps the apparent movement of these intracellular structures -- indirectly but systematically. Rapidly fragmented mitochondria create more active, widespread speckle movement; slow, interconnected mitochondrial networks create more localized, gradual shifts.
The TV-L1 optical flow algorithm used here computes this motion field by solving an optimization problem that balances fidelity to the observed intensity changes (the data term) with smoothness of the motion field (the regularization term). The L1 norm in the data term makes it more robust to noise and large discontinuities compared to the classical Lucas-Kanade optical flow, which is important in DLS videos where speckle patterns can change abruptly due to random scattering events.
By stacking the horizontal and vertical components of optical flow from consecutive frames along the channel dimension of the input tensor, the temporal stream receives a spatial map of motion that is both interpretable and information-rich. This input format allows ResNet-50 -- originally designed for static image recognition -- to effectively process video dynamics without requiring architectural changes for temporal modeling.
This study introduces a conceptually novel approach to cancer subtyping: rather than measuring molecular markers (receptor proteins, DNA, RNA) or analyzing tissue morphology in stained slides, it reads the physical dynamics of living cells. The biological differences between TNBC and other subtypes -- particularly the higher mitochondrial fission activity -- are not just molecular signatures; they produce measurable changes in how cells scatter and move light. This physical readout requires no chemical labeling, no tissue processing, and no molecular assay infrastructure.
The approach is particularly relevant for TNBC because mitochondrial fission is not merely a correlate of TNBC aggressiveness -- it is an active driver of it. The elevated activity of Drp1 (the fission-promoting protein) in TNBC cells actively promotes tumor cell survival and invasion. This means the DLS signal is not an incidental byproduct but a direct measurement of a biologically meaningful process. The same mitochondrial dynamics that make TNBC dangerous are also what make it detectable by this method.
A key practical advantage is the speed and simplicity of the measurement. Each cell video takes approximately 10 seconds, and the entire analysis pipeline from data collection to classification is automated. No trained pathologist, no staining protocol, and no IHC testing equipment are required. For research settings where rapid, high-throughput cell characterization is needed -- such as drug screening or cell line quality control -- this represents a genuinely useful capability.
Important limitations apply. This proof-of-concept was conducted on cultured cell lines, not patient tumor samples. Cell lines grown in the lab behave differently from cells in an actual tumor, which exist in a complex tissue environment with stroma, immune cells, and extracellular matrix. Whether the DLS signatures observed here will translate to primary tumor cells obtained from biopsies remains to be established. The current system also classifies only two subtypes (TNBC vs. HER2-positive); extending it to all four breast cancer molecular subtypes would require substantially more training data.
This study demonstrated a proof of concept for label-free TNBC identification using dynamic light scattering microscopy combined with a two-stream deep learning classifier. The approach achieved an AUC of 0.95 and accuracy of 89% for distinguishing TNBC cells from HER2-positive cells using only 10-second videos of unstained, living cells in suspension.
The mechanistic chain linking biology to signal to classification is well-supported. TNBC cells have more fragmented mitochondria (confirmed by confocal imaging), which produce more active intracellular motion (confirmed by autocorrelation analysis), which creates faster-fluctuating DLS speckle patterns (confirmed by mitochondrial inhibition experiments), which the temporal stream of the deep learning model captures through optical flow. Each link in this chain was experimentally verified rather than assumed.
Future directions include testing on primary cell samples from patient biopsies, expanding the classification to all four breast cancer molecular subtypes, miniaturizing the optical setup for potential clinical or point-of-care applications, and exploring whether DLS dynamics can predict treatment response to mitochondria-targeting therapies -- a logical next step given the central role mitochondrial fission plays in TNBC cell survival and metastasis.