TL;DR
When modeling tumor growth, the standard practice of assuming constant measurement error is often wrong. This paper compares five different "likelihood" models for the Gompertz growth curve and finds that accounting for errors that scale with tumor volume—specifically using a threshold-based error model—provides a significantly more accurate fit and offers new clinical insights into tumor morphology.
What Problem Does This Paper Address?
In the world of oncology, the Gompertz model is a classic mathematical tool used to describe how solid tumors grow. It captures the "S-curve" behavior where growth starts fast and slows down as the tumor reaches a carrying capacity.
However, there is a hidden challenge: measurement error. As a tumor gets larger, its shape often shifts from a neat sphere to an irregular ellipsoid, and its surface becomes "bumpy." This makes it much harder to measure accurately. Most researchers ignore this and use a standard "least-squares" approach, which assumes the measurement error is the same regardless of whether the tumor is tiny or huge (constant variance).
The authors ask a critical question: Should we stick to the simple, standard error models, or do we need more complex "likelihood functions" that account for these escalating errors?
Methods
The researchers compared five distinct error models (likelihood functions) to see which best fits real-world tumor data:
- Norm: The standard Normal distribution with constant error.
- Prop: A Normal distribution where error is proportional to tumor volume.
- Thres: A "Threshold" model where error is constant until a certain volume ($V_m$), then becomes proportional.
- Stud: A Student-t distribution (more robust to outliers) with constant error.
- Stud Prop: A Student-t distribution with volume-proportional error.
To evaluate these, they used a Bayesian framework, employing tools like the Bayesian Information Criterion (BIC), Deviance Information Criterion (DIC), and Bayes Factors. They also used the Kolmogorov-Smirnov (KS) test on the residuals to see if the chosen error model actually matched the statistical "noise" in the data.
Datasets
The study utilized experimental volume data from three types of untreated tumors in mice:
- Ehrlich tumors.
- Fibrosarcoma Sa-37 tumors.
- F3II mammary carcinoma (with separate datasets for male and female mice).
These datasets provided a robust testing ground for the models across different histological types and biological contexts.
Key Insights and Ideas
- The "Thres" Model is a Winner: The threshold model often provided the best balance. It suggests that measurement error is stable while the tumor is small and spherical but begins to escalate once the tumor hits a specific size.
- Biological Interpretation of Math: The threshold parameter ($V_m$) isn't just a number; the authors suggest it could represent the transition from the avascular phase (diffusion-dominated) to the vascular phase (angiogenesis-dominated), where the tumor's shape becomes more irregular.
- Constant Variance is Statistically Flawed: Models assuming constant error (Norm and Stud) frequently failed the KS tests. This means the "noise" in the data isn't actually constant, and using these models might lead to biased parameter estimates.
- The "True" Volume Bias: Interestingly, models with variable dispersion suggest that the actual tumor volume is likely higher than what is typically measured, as measurement techniques tend to underestimate large, irregular masses.
Key Results
- Quantitative Superiority: Models with volume-dependent dispersion (Prop, Thres, Stud Prop) consistently outperformed constant-variance models in DIC and KS tests.
- Residual Analysis: In the Ehrlich tumor dataset, the "Thres" model achieved the highest fraction of non-rejected normality tests ($fr(KS) \approx 0.275$), while the standard "Norm" model was near zero ($2.0 \times 10^{-4}$).
- BIC vs. DIC: While BIC (which heavily penalizes complexity) sometimes favored simpler models like "Norm," the DIC and Bayes Factors generally pointed toward the more complex, volume-dependent models as being more representative of the underlying biology.
Limitations and Future Work
- Computational Complexity: Moving away from simple least-squares requires more advanced algorithms (like Nelder-Mead or MCMC), which might be a barrier for some clinical researchers.
- Selection of Training Fractions: The use of Fractional Bayes Factors (FBF) requires choosing a training fraction ($b$), which the authors noted can introduce its own bias if not handled carefully.
- Untreated Focus: This study looked at untreated tumors. Future work is needed to see how these likelihood choices hold up when a tumor is shrinking or changing shape due to chemotherapy or radiation.
- Clinical Integration: The authors recommend that while the "Norm" model should remain a baseline for compatibility with past research, at least one variable-dispersion model should be included in future studies to provide a more realistic estimate of tumor size.
Summary updated